Counting 3,663 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Author
Contributors
Project Page
https://rubixml.com/
Last Commit
Mar. 23, 2019
Created
Feb. 14, 2018

Rubix ML for PHP

PHP from Packagist Latest Stable Version Downloads from Packagist Travis GitHub license

A high-level machine learning library that allows you to build programs that learn from data using the PHP language.

  • Easy and fast prototyping with user-friendly API
  • 40+ modern supervised and unsupervised learners
  • Modular architecture combines power and flexibility
  • Open source and free to use commercially

Installation

Install Rubix ML using Composer:

$ composer require rubix/ml

Requirements

  • PHP 7.1.3 or above

Optional

Documentation

Table of Contents


Basic Introduction

Machine learning is the process by which a computer program is able to progressively improve performance on a certain task through training and data without explicitly being programmed. There are two types of machine learning that Rubix supports out of the box - Supervised and Unsupervised.

  • Supervised learning is a technique that uses a labeled dataset in which the outcome of each sample has been labeled by a human expert prior to training. There are two types of supervised learning to consider in Rubix:
    • Classification is the problem of identifying which class a particular sample belongs to. For example, one task may be in determining a particular species of flower or predicting someone's MBTI personality type.
    • Regression aims at predicting continuous values such as the sale price of a house or the position (in degrees) of the steering wheel of an automobile. The major difference between classification and regression is that, while there are a finite number of classes that a sample can belong to, there are infinitely many real (continuous) values that are possible to predict.
  • Unsupervised learning by contrast does not use a labeled dataset. Instead, it focuses on finding patterns within the raw samples.
    • Clustering is the grouping of data points in such a way that members of the same group are more similar (homogeneous) than the rest of the samples. You can think of clustering as assigning a class label to an otherwise unlabeled sample. An example where clustering is used is in differentiating PET scan tissues or segmenting a customer base.
    • Anomaly Detection is the process of flagging samples that appear to be generated from a mechanism other than one that produces nominal data. Anomalous samples can indicate adversarial activity or exceptional circumstances such as fraud or a cyber attack.
    • Dimensionality Reduction is used in visualizing high dimensional datasets, embedding sparse feature representations, and reducing model size by producing a low dimensional representation of the original feature space.

Obtaining Data

Machine learning projects typically begin with a question. For example, you might want to answer the question "who of my friends are most likely to stay married to their spouse?" One way to go about answering this question with machine learning would be to go out and ask a bunch of happily married and divorced couples the same set of questions about their partner and then use that data to build a model of what a successful marriage looks like. Later, you can use that model to make predictions based on the answers you get from your friends. Specifically, the answers you collect are called features and they constitute measurements of some phenomena being observed. The number of features in a sample is called the dimensionality of the sample. For example, a sample with 20 features is said to be 20 dimensional.

An alternative to collecting data yourself is to download one of the many open datasets that are free to use from a public repository. The advantages of using a public dataset is that, usually, the data has already been cleaned and prepared for you. We recommend the University of California Irvine Machine Learning Repository as a great place to get started with using open source datasets.

There are many PHP libraries that help make extracting data from various sources such as CSV, database, and the cloud easy and intuitive, and we recommend checking those out as well.

Here are some libraries that we recommend for data extraction:

The Dataset Object

In Rubix, data is passed around in specialized data containers called Datasets. Dataset objects internally handle selecting, splitting, folding, transforming, and randomizing the samples and labels contained within. In general, there are two types of datasets, Labeled and Unlabeled. Labeled datasets are used for supervised learning and for providing the ground-truth during cross validation. Unlabeled datasets are used for unsupervised learning and for making predictions (which we call inference) on unknown samples.

For the following example, suppose that you went out and asked 100 couples (50 married and 50 divorced) about their partner's communication skills (between 1 and 5), attractiveness (between 1 and 5), and time spent together per week (hours per week). You could construct a Labeled Dataset from this data like so:

use Rubix\ML\Datasets\Labeled;

$samples = [
    [3, 4, 50.5], [1, 5, 24.7], [4, 4, 62.0], [3, 2, 31.1], ...
];

$labels = ['married', 'divorced', 'married', 'divorced', ...];

$dataset = new Labeled($samples, $labels);

Choosing an Estimator

Estimators make up the core of the Rubix library as they are responsible for making predictions. There are many different algorithms to choose from and each one performs differently. Choosing the right Estimator for the job is crucial to creating a system that balances accuracy and performance.

In practice, you will test out a number of different estimators to get the best sense of what works for your particular dataset. However, for our example problem we will just focus on a simple classifier called K Nearest Neighbors. Since the label of each training sample we collect will be a discrete class (married couples or divorced couples), we need an Estimator that is designed to output class predictions. The K Nearest Neighbors classifier works by locating the closest training samples to an unknown sample and choosing the class label that appears most often.

Creating the Estimator Instance

Like most Estimators, the K Nearest Neighbors (KNN) classifier requires a set of parameters (called hyper-parameters) to be chosen up front by the user. These parameters can be selected based on some prior knowledge of the problem space, or at random. The defaults provided in Rubix are a good place to start for most machine learning problems. In addition, the library provides a meta-Estimator called Grid Search that searches for the best combination of hyper-parmeters for a particular estimator given a set of possible values. For the purposes of our simple example we will just go with our intuition and choose the parameters outright.

Note: You can find a full description of all of the K Nearest Neighbors hyper-parameters in the API reference guide.

In KNN, the hyper-parameter k is the number of nearest points from the training set to compare an unknown sample to in order to infer its class label. For example, if the 5 closest neighbors to a given unknown sample have 4 married labels and 1 divorced label, then the algorithm will output a prediction of married with a probability of 0.8.

The second hyper-parameter is the distance kernel that determines how distance is measured within the model. We'll go with standard Euclidean distance for now.

To instantiate a K Nearest Neighbors classifier:

use Rubix\ML\Classifiers\KNearestNeighbors;
use Rubix\ML\Kernels\Distance\Euclidean;

// Using the default hyper-parameters
$estimator = new KNearestNeighbors();

// Specifying the hyper-parameters
$estimator = new KNearestNeighbors(5, new Euclidean());

Now that we've chosen and instantiated our estimator and our Dataset object is ready to go, we are ready to train the model and use it to make some predictions.

Training and Prediction

Training is the process of feeding the learning algorithm data so that it can build a model of the problem. A trained model consists of all of the parameters (except hyper-parameters) that are required for the estimator to make predictions. If you try to make predictions using an untrained learner, it will throw an exception.

Passing the Labeled dataset to the instantiated learner, we can train our K Nearest Neighbors classifier like so:

$estimator->train($dataset);

We can verify that the learner has been trained by calling the trained() method:

var_dump($estimator->trained());

Output:

bool(true)

For our 100 sample example training set, training should only take a matter of microseconds, but larger datasets with higher dimensionality and fancier learning algorithms can take much longer. Once the estimator has been fully trained, we can now feed in some unknown samples to see what the model predicts.

Turning back to our example problem, suppose that we went out and collected 5 new data points from our friends using the same questions we asked the couples we interviewed for our training set. We could make predictions on whether they will stay married or get divorced by taking their answers as features and running them in an Unlabeled dataset through the trained Estimator's predict() method.

use Rubix\ML\Dataset\Unlabeled;

$unknown = [
    [4, 3, 44.2], [2, 2, 16.7], [2, 4, 19.5], [1, 5, 8.6], [3, 3, 55.0],
];

$dataset = new Unlabeled($unknown);

$predictions = $estimator->predict($dataset);

var_dump($predictions);

Output:

array(5) {
	[0] => 'married'
	[1] => 'divorced'
	[2] => 'divorced'
	[3] => 'divorced'
	[4] => 'married'
}

Evaluating Model Performance

Making predictions is not very useful unless the estimator can correctly generalize what it has learned during training to the real world. Cross Validation is a process by which we can test the model for its generalization ability. For the purposes of this introduction, we will use a simple form of cross validation called Hold Out. The Hold Out validator will take care of splitting the dataset into training and testing sets automatically, such that a portion of the data is held out to be used for testing (or validating) the model. The reason we do not use all of the data for training is because we want to test the Estimator on samples that it has never seen before.

The Hold Out validator requires you to set the ratio of testing to training samples as a constructor parameter. In this case, let's choose to use a factor of 0.2 (20%) of the dataset for testing leaving the rest (80%) for training. Typically, 0.2 is a good default choice however your mileage may vary. The important thing to note here is the trade off between more data for training and more data to produce precise testing results. Once you get the hang of Hold Out, the next step is to consider more advanced cross validation techniques such as K Fold, Leave P Out, and Monte Carlo simulations.

To return a score from the Hold Out validator using the Accuracy metric just pass it the untrained estimator instance and a dataset:

use Rubix\ML\CrossValidation\HoldOut;
use Rubix\ML\CrossValidation\Metrics\Accuracy;

$validator = new HoldOut(0.2);

$score = $validator->test($estimator, $dataset, new Accuracy());

var_dump($score);

Output:

float(0.945)

Visualization

Visualization is how you communicate the findings of your experiment to the end-user and is key to deriving value from your hard work. Although visualization is important (important enough for us to mention it), we consider it to be beyond the scope of Rubix . Therefore, we leave you with the freedom to chose one of the many great plotting and visualization applications out there to communicate the insights you obtain using Rubix.

If you are just looking for a quick way to visualize data then we recommend exporting it to a file (JSON or CSV for example) and importing it into your favorite plotting or spreadsheet software such as Plotly, Tableu, Google Sheets, or Excel. PHP has built in functions for manipulating both JSON and CSV formats, and there are a number of libraries available that help reading and writing these formats to file from PHP.

If you are looking to publish your visualizations to the world, we highly recommend D3.js since it is an amazing data-driven framework written in Javascript that plays well with PHP.

Next Steps

After you've gone through this basic introduction to machine learning, we highly recommend checking out all of the example projects and reading over the API Reference to get a powerful understanding for what you can do with Rubix. If you have a question or need help, feel free to post on our Github page. We'd love to hear from you.


System Architecture

The Rubix architecture is defined by a few key abstractions and their corresponding types and interfaces.

Rubix ML System Architecture


API Reference

This section breaks down the public application programming interface (API) of each Rubix ML component in detail.

Dataset Objects

In Rubix, data is passed around using specialized data structures called Dataset objects. Dataset objects can hold a heterogeneous mix of categorical and continuous data and make it easy to transport data in a canonical way.

Note: There are two types of features that estimators can process i.e categorical and continuous. Any numerical (integer or float) datum is considered continuous and any string datum is considered categorical by convention throughout Rubix.

The Dataset interface has a robust API designed to make working on datasets fast and easy. Below you'll find a description of the various methods available on the basic interface.

Stacking

Stack a number of dataset objects on top of each other and return a single dataset:

public static stack(array $datasets) : self;

Selecting

Return all the samples in the dataset:

public samples() : array

Select the sample at row offset:

public row(int $index) : array

Select the values of a feature column at offset:

public column(int $index) : array

Return the first n rows of data in a new dataset object:

public head(int $n = 10) : self

Return the last n rows of data in a new dataset object:

public tail(int $n = 10) : self

Example:

// Return the sample matrix
$samples = $dataset->samples();

// Return just the first 5 rows in a new dataset
$subset = $dataset->head(5);

Properties

Return the number of rows in the dataset:

public numRows() : int

Return the number of columns in the dataset:

public numColumns() : int

Return the integer encoded column types for each feature column:

public types() : array

Return the integer encoded column type given a column index:

public columnType(int $index) : int

Return the range for each feature column:

public ranges() : array

Return the range of a feature column. The range for a continuous column is defined as the minimum and maximum values, and for categorical columns the range is defined as every unique category.

public columnRange(int $index) : array

Splitting, Folding, and Batching

Remove n rows from the dataset and return them in a new dataset:

public take(int $n = 1) : self

Leave n samples on the dataset and return the rest in a new dataset:

public leave(int $n = 1) : self

Split the dataset into left and right subsets given by a ratio:

public split(float $ratio = 0.5) : array

Partition the dataset into left and right subsets based on the value of a feature in a specified column:

public partition(int $index, mixed $value) : array

Fold the dataset k - 1 times to form k equal size datasets:

public fold(int $k = 10) : array

Batch the dataset into subsets of n rows per batch:

public batch(int $n = 50) : array

Example:

// Remove the first 5 rows and return them in a new dataset
$subset = $dataset->take(5);

// Split the dataset into left and right subsets
[$left, $right] = $dataset->split(0.5);

// Partition the dataset by the feature column at index 4 by value '50'
[$left, $right] = $dataset->partition(4, 50);

// Fold the dataset into 8 equal size datasets
$folds = $dataset->fold(8);

Randomizing

Randomize the order of the Dataset and return it:

public randomize() : self

Generate a random subset with replacement of size n:

public randomSubsetWithReplacement($n) : self

Generate a random weighted subset with replacement of size n:

public randomWeightedSubsetWithReplacement($n, array $weights) : self

Example:

// Randomize and split the dataset into two subsets
[$left, $right] = $dataset->randomize()->split(0.8);

// Generate a bootstrap dataset of 500 random samples
$subset = $dataset->randomSubsetWithReplacement(500);

Filtering

To filter a Dataset by a feature column:

public filterByColumn(int $index, callable $fn) : self

Example:

$tallPeople = $dataset->filterByColumn(2, function ($value) {
	return $value > 178.5;
});

Sorting

To sort a dataset in place by a specific feature column:

public sortByColumn(int $index, bool $descending = false) : self

Example:

var_dump($dataset->samples());

$dataset->sortByColumn(2, false);

var_dump($dataset->samples());

Output:

array(3) {
    [0]=> array(3) {
	    [0]=> string(4) "mean"
	    [1]=> string(4) "furry"
	    [2]=> int(8)
    }
    [1]=> array(3) {
	    [0]=> string(4) "nice"
	    [1]=> string(4) "rough"
	    [2]=> int(1)
    }
    [2]=> array(3) {
	    [0]=> string(4) "nice"
	    [1]=> string(4) "rough"
	    [2]=> int(6)
    }
}

array(3) {
    [0]=> array(3) {
	    [0]=> string(4) "nice"
	    [1]=> string(4) "rough"
	    [2]=> int(1)
    }
    [1]=> array(3) {
	    [0]=> string(4) "nice"
	    [1]=> string(4) "rough"
	    [2]=> int(6)
    }
    [2]=> array(3) {
	    [0]=> string(4) "mean"
	    [1]=> string(4) "furry"
	    [2]=> int(8)
    }
}

Prepending and Appending

To prepend a given dataset onto the beginning of another dataset:

public prepend(Dataset $dataset) : self

To append a given dataset onto the end of another dataset:

public append(Dataset $dataset) : self

Applying a Transformation

You can apply a fitted Transformer to a Dataset directly passing it to the apply method on the Dataset.

public apply(Transformer $transformer) : void

Example:

use Rubix\ML\Transformers\OneHotEncoder;

$transformer = new OneHotEncoder();

$transformer->fit($dataset);

$dataset->apply($transformer);

Labeled

For supervised Estimators you will need to train it with a Labeled dataset consisting of samples with the addition of labels that correspond to the observed outcome of each sample. Splitting, folding, randomizing, sorting, and subsampling are all done while keeping the indices of samples and labels aligned. In addition to the basic Dataset interface, the Labeled class can sort and stratify the data by label as well.

Parameters:

# Param Default Type Description
1 samples array A 2-dimensional array consisting of rows of samples and columns with feature values.
2 labels array A 1-dimensional array of labels that correspond to the samples in the dataset.
3 validate true bool Should we validate the data?

Additional Methods:

Build a new labeled dataset with validation:

public static build(array $samples = [], array $labels = []) : self

Build a new labeled dataset foregoing validation:

public static quick(array $samples = [], array $labels = []) : self

Build a dataset using a pair of iterators:

public static fromIterator(iterable $samples, iterable $labels) : self

Return an array of labels:

public labels() : array

Return the samples and labels in a single array:

public zip() : array

Return the label at the given row offset:

public label(int $index) : mixed

Return the type of the label encoded as an integer:

public labelType() : int

Return all of the possible outcomes i.e. the unique labels:

public possibleOutcomes() : array

Transform the labels in the dataset using a function:

public transformLabels(callable $fn) : void

The function is given a label as its only argument and should return the new label as a continuous or categorical value.

$dataset->transformLabels(function ($label) {
	return $label === 1 ? 'female' : 'male';
});

Filter the dataset by label:

public filterByLabel(callable $fn) : self

Sort the dataset by label:

public sortByLabel(bool $descending = false) : self

Group the samples by label and return them in their own dataset:

public stratify() : array

Split the dataset into left and right stratified subsets with a given ratio of samples in each:

public stratifiedSplit($ratio = 0.5) : array

Return k equal size subsets of the dataset:

public stratifiedFold($k = 10) : array

Example:

use Rubix\ML\Datasets\Labeled;

// Import samples and labels

$dataset = Labeled::build($samples, $labels);  // Build a new dataset with validation

// or ...

$dataset = Labeled::quick($samples, $labels);  // Build a new dataset without validation

// or ...

$dataset = new Labeled($samples, $labels, true);  // Use the full constructor

// Transform integer encoded labels to strings
$dataset->transformLabels(function ($label) {
	switch ($label) {
		case 1:
			return 'male';
		
		case 2:
			return 'female';

		default:
			return 'unknown';
	}
});

// Return all the labels in the dataset
$labels = $dataset->labels();

// Return the label at offset 3
$label = $dataset->label(3);

// Return all possible unique labels
$outcomes = $dataset->possibleOutcomes();

var_dump($labels);
var_dump($label);
var_dump($outcomes);

Output:

array(4) {
    [0]=> string(5) "female"
    [1]=> string(4) "male"
    [2]=> string(5) "female"
    [3]=> string(4) "male"
}

string(4) "male"

array(2) {
	[0]=> string(5) "female"
	[1]=> string(4) "male"
}

Example:

// Fold the dataset into 5 equal size stratified subsets
$folds = $dataset->stratifiedFold(5);

// Split the dataset into two stratified subsets
[$left, $right] = $dataset->stratifiedSplit(0.8);

// Put each sample with label x into its own dataset
$strata = $dataset->stratify();

Unlabeled

Unlabeled datasets can be used to train unsupervised Estimators and for feeding data into an Estimator to make predictions.

Parameters:

# Param Default Type Description
1 samples array A 2-dimensional array consisting of rows of samples and columns with feature values.
2 validate true bool Should we validate the input?

Additional Methods:

Build a new unlabeled dataset with validation:

public static build(array $samples = []) : self

Build a new unlabeled dataset foregoing validation:

public static quick(array $samples = []) : self

Build a dataset with an iterator:

public static fromIterator(iterable $samples) : self

Example:

use Rubix\ML\Datasets\Unlabeled;

$dataset = Unlabeled::build($samples);  // Build a new dataset with validation

// or ...

$dataset = Unlabeled::quick($samples);  // Build a new dataset without validation

// or ...

$dataset = new Unlabeled($samples, true);  // Use the full constructor

Estimators

Estimators consist of various Classifiers, Regressors, Clusterers, Embedders, and Anomaly Detectors that make predictions based on their training. Estimators that can be trained using data are called Learners and they can either be supervised or unsupervised depending on the task. Estimators can employ methods on top of the basic API by implementing a number of addon interfaces such as Online, Probabilistic, Persistable, and Verbose. The most basic Estimator is one that outputs an array of predictions given a dataset of unknown or testing samples.

To make predictions, pass the estimator a dataset object filled with samples you'd like to predict:

public predict(Dataset $dataset) : array

Note: The return value of predict() is an array containing the predictions indexed in the same order that they were fed into the estimator.

Learner

Most estimators have the ability to be trained with data. These estimators are called Learners and require training before they are able make predictions. Training is the process of feeding data to the learner so that it can formulate a generalized function that maps future samples to good predictions.

To train an learner pass it a training dataset:

public train(Dataset $training) : void

Return whether or not the learner has been trained:

public trained() : bool

Example:

$estimator->train($dataset);

Note: Calling train() on an already trained estimator will cause any previous training to be lost. If you would like to be able to train a model incrementally, see the Online Estimator interface.

Online

Certain estimators that implement the Online interface can be trained in batches. Estimators of this type are great for when you either have a continuous stream of data or a dataset that is too large to fit into memory. Partial training allows the model to evolve as new information about the world is acquired.

You can partially train an online estimator by:

public partial(Dataset $dataset) : void

Example:

$folds = $dataset->fold(3);

$estimator->train($folds[0]);

$estimator->partial($folds[1]);

$estimator->partial($folds[2]);

Note: an Estimator will continue to train as long as you are using the partial() method, however, calling train() on a trained or partially trained Estimator will reset it back to baseline first.


Probabilistic

Estimators that implement the Probabilistic interface have an additional method that returns an array of probability scores of each possible class, cluster, etc. Probabilities are useful for ascertaining the degree to which the estimator is certain about a particular prediction.

Return the probability estimates of a prediction:

public proba(Dataset $dataset) : array

Example:

$probabilities = $estimator->proba($dataset);  

var_dump($probabilities);

Output:

array(2) {
	[0] => array(2) {
		['married'] => 0.975,
		['divorced'] => 0.025,
	}
	[1] => array(2) {
		['married'] => 0.200,
		['divorced'] => 0.800,
	}
}

Ranking

In the way that a Probabilistic estimator ranks the outcome of a particlar sample as a normalized (between 0 and 1) value, Ranking estimators rank the outcome by an arbitrary score. The purpose of a Ranking estimator is so that you are able to sort the samples by the output. This is useful in cases such as Anomaly Detection where an analyst can flag the top n outliers by rank for further investigation.

To rank the dataset by an artitrary scoring function:

public rank(Dataset $dataset) : array

Example:

$scores = $estimator->rank($dataset);

var_dump($scores);

Output:

array(3) {
  [0]=> float(1.80)
  [1]=> int(1.25)
  [2]=> int(9.45)
}

Verbose

Verbose estimators are capable of logging errors and important training events to any PSR-3 compatible logger such as Monolog, Analog, or the included Screen Logger. Logging is especially useful for monitoring the progress of the underlying learning algorithm in real time.

To set the logger pass in any PSR-3 compatible logger instance:

public setLogger(LoggerInterface $logger) : void

Example:

use Rubix\ML\Other\Loggers\Screen;

$estimator->setLogger(new Screen('sentiment'));

Anomaly Detectors

Anomaly detection is the process of identifying samples that do not conform to an expected pattern. The output prediction of a detector is a binary encoding (either 0 for a normal sample or 1 for a detected anomaly).

Isolation Forest

An ensemble detector comprised of Isolation Trees each trained on a different subset of the training set. The Isolation Forest works by averaging the isolation score of a sample across a user-specified number of trees.

Interfaces: Learner, Ranking, Persistable
Compatibility: Categorical, Continuous

Parameters:

# Param Default Type Description
1 estimators 300 int The number of estimators to train in the ensemble.
2 contamination 0.1 float The percentage of outliers that are assumed to be present in the training set.
3 ratio 0.2 float The ratio of random samples to train each estimator with.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\AnomalyDetection\IsolationForest;

$estimator = new IsolationForest(300, 0.01, 0.2);

References:

  • F. T. Liu et al. (2008). Isolation Forest.
  • F. T. Liu et al. (2011). Isolation-based Anomaly Detection.

K-d LOF

A k-d tree accelerated version of Local Outlier Factor which benefits from fast nearest neighbors search.

Interfaces: Learner, Ranking, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 k 20 int The k nearest neighbors that form a local region.
2 contamination 0.1 float The percentage of outliers that are assumed to be present in the training set.
3 kernel Euclidean object The distance kernel used to compute the distance between sample points.
4 max leaf size 30 int The max number of samples in a leaf node (neighborhood).

Additional Methods:

Return the height of the tree:

public height() : int

Return the balance of the tree:

public balance() : int

Example:

use Rubix\ML\AnomalyDetection\KDLOF;
use Rubix\ML\Kernels\Distance\Euclidean;

$estimator = new KDLOF(20, 0.1, new Euclidean(), 30);

References:

  • M. M. Breunig et al. (2000). LOF: Identifying Density-Based Local Outliers.

Local Outlier Factor

Local Outlier Factor (LOF) measures the local deviation of density of a given sample with respect to its k nearest neighbors. As such, LOF only considers the local region (or neighborhood) of an unknown sample which enables it to detect anomalies within individual clusters of data.

Interfaces: Learner, Online, Ranking, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 k 20 int The k nearest neighbors that form a local region.
2 contamination 0.1 float The percentage of outliers that are assumed to be present in the training set.
3 kernel Euclidean object The distance kernel used to compute the distance between sample points.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\AnomalyDetection\LocalOutlierFactor;
use Rubix\ML\Kernels\Distance\Minkowski;

$estimator = new LocalOutlierFactor(20, 0.1, new Minkowski(3.5));

References:

  • M. M. Breunig et al. (2000). LOF: Identifying Density-Based Local Outliers.

LODA

Lightweight Online Detector of Anomalies uses sparse random projection vectors to produce an ensemble of unique one dimensional equi-width histograms able to estimate the probability density of an unknown sample. The anomaly score is given by the negative log likelihood whose upper threshold can be set by the user.

Interfaces: Learner, Online, Ranking, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 threshold 5.5 float The threshold anomaly score to be flagged as an outlier.
2 k 100 int The number of random projections and histograms.
3 bins Auto int The number of bins for each equi-width histogram.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\AnomalyDetection\LODA;

$estimator = new LODA(3.5, 250, 6);

References:

  • T. Pevný. (2015). Loda: Lightweight on-line detector of anamalies.
  • L. Birg´e et al. (2005). How Many Bins Should Be Put In A Regular Histogram.

One Class SVM

An unsupervised Support Vector Machine used for anomaly detection. The One Class SVM aims to find a maximum margin between a set of data points and the origin, rather than between classes such as with multiclass SVM (SVC).

Note: This estimator requires the SVM PHP extension which uses the LIBSVM engine written in C++ under the hood.

Interfaces: Learner, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 nu 0.1 float An upper bound on the percentage of margin errors and a lower bound on the percentage of support vectors.
2 kernel RBF object The kernel function used to express non-linear data in higher dimensions.
3 shrinking true bool Should we use the shrinking heuristic?
4 tolerance 1e-3 float The minimum change in the cost function necessary to continue training.
5 cache size 100. float The size of the kernel cache in MB.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\AnomalyDetection\OneClassSVM;
use Rubix\ML\Kernels\SVM\Polynomial;

$estimator = new OneClassSVM(0.1, new Polynomial(4), true, 1e-3, 100.);

References:

  • C. Chang et al. (2011). LIBSVM: A library for support vector machines.

Robust Z Score

A quick global anomaly detector that uses a modified Z score which is robust to outliers to detect anomalies within a dataset. The modified Z score consists of taking the median and median absolute deviation (MAD) instead of the mean and standard deviation (standard Z score) thus making the statistic more robust to training sets that may already contain outliers. Outliers can be flagged in one of two ways. First, their average Z score can be above the user-defined tolerance level or an individual feature's score could be above the threshold (hard limit).

Interfaces: Learner, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 tolerance 3.0 float The average z score to tolerate before a sample is considered an outlier.
2 threshold 3.5 float The threshold z score of a individual feature to consider the entire sample an outlier.

Additional Methods:

Return the median of each feature column in the training set:

public medians() : ?array

Return the median absolute deviation (MAD) of each feature column in the training set:

public mads() : ?array

Example:

use Rubix\ML\AnomalyDetection\RobustZScore;

$estimator = new RobustZScore(1.5, 3.0);

References:

  • P. J. Rousseeuw et al. (2017). Anomaly Detection by Robust Statistics.

Classifiers

Classifiers are a type of estimator that predict discrete outcomes such as categorical class labels.

AdaBoost

Short for Adaptive Boosting, this ensemble classifier can improve the performance of an otherwise weak classifier by focusing more attention on samples that are harder to classify.

Note: The default base classifier is a Decision Stump i.e a Classification Tree with a max depth of 1.

Interfaces: Learner, Probabilistic, Verbose, Persistable
Compatibility: Depends on base learner

Parameters:

# Param Default Type Description
1 base Classification Tree object The base weak classifier to be boosted.
2 estimators 100 int The number of estimators to train in the ensemble.
3 rate 1.0 float The learning rate i.e step size.
4 ratio 0.8 float The ratio of samples to subsample from the training set per epoch.
5 tolerance 1e-3 float The amount of validation error to tolerate before an early stop is considered.

Additional Methods:

Return the calculated weight values of the last trained dataset:

public weights() : array

Return the influence scores for each boosted classifier:

public influences() : array

Return the training error at each epoch:

public steps() : array

Example:

use Rubix\ML\Classifiers\AdaBoost;
use Rubix\ML\Classifiers\ExtraTreeClassifier;

$estimator = new AdaBoost(new ExtraTreeClassifier(3), 100, 0.1, 0.5, 1e-2);

References:

  • Y. Freund et al. (1996). A Decision-theoretic Generalization of On-line Learning and an Application to Boosting.
  • J. Zhu et al. (2006). Multi-class AdaBoost.

Classification Tree

A binary tree-based classifier that minimizes gini impurity to greedily construct a decision tree for classification.

Interfaces: Learner, Probabilistic, Verbose, Persistable
Compatibility: Categorical, Continuous

Parameters:

# Param Default Type Description
1 max depth PHP_INT_MAX int The maximum depth of a branch.
2 max leaf size 3 int The max number of samples that a leaf node can contain.
3 min purity increase 0. float The minimum increase in purity necessary for a node not to be post pruned.
4 max features Auto int The max number of features to consider when determining a best split.
5 tolerance 1e-3 float A small amount of impurity to tolerate when choosing a best split.

Additional Methods:

Return the feature importances calculated during training indexed by feature column:

public featureImportances() : array

Return the height of the tree:

public height() : int

Return the balance of the tree:

public

Example:

use Rubix\ML\Classifiers\ClassificationTree;

$estimator = new ClassificationTree(30, 7, 0.1, 4, 1e-4);

Committee Machine

A voting ensemble that aggregates the predictions of a committee of heterogeneous classifiers (referred to as experts). The committee employs a user-specified influence-based scheme to make final predictions.

Note: Influence values can be arbitrary as they are normalized upon instantiation anyways.

Interfaces: Learner, Ensemble, Probabilistic, Persistable
Compatibility: Depends on base learners

Parameters:

# Param Default Type Description
1 experts array An array of classifier instances that comprise the committee.
2 influences 1 / n array The influence values of each expert in the committee.

Additional Methods:

Return the normalized influence scores of each estimator in the committee:

public influences() : array

Example:

use Rubix\ML\Classifiers\CommitteeMachine;
use Rubix\ML\Classifiers\RandomForest;
use Rubix\ML\Classifiers\ClassificationTree;
use Rubix\ML\Classifiers\SoftmaxClassifier;
use Rubix\ML\NeuralNet\Optimizers\Adam;
use Rubix\ML\Classifiers\KNearestNeighbors;

$estimator = new CommitteeMachine([
	new SoftmaxClassifier(100, new Adam(0.001)),
	new RandomForest(new ClassificationTree(4), 100, 0.3),
	new KNearestNeighbors(3),
], [
	4, 6, 5, // Arbitrary influence values for each expert
]);

Dummy Classifier

A classifier that uses a user-defined Guessing Strategy to make predictions. Dummy Classifier is useful to provide a sanity check and to compare performance with an actual classifier.

Interfaces: Learner, Persistable
Compatibility: Categorical, Continuous, Resource

Parameters:

# Param Default Type Description
1 strategy Popularity Contest object The guessing strategy to employ when guessing the outcome of a sample.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\Classifiers\DummyClassifier;
use Rubix\ML\Other\Strategies\PopularityContest;

$estimator = new DummyClassifier(new PopularityContest());

Extra Tree Classifier

An Extremely Randomized Classification Tree - these trees differ from standard Classification Trees in that they choose the best split drawn from a random set determined by max features, rather than searching the entire column. Extra Trees work well in ensembles such as Random Forest or AdaBoost as the weak learner or they can be used on their own. The strength of Extra Trees are computational efficiency as well as increasing variance of the prediction (if that is desired).

Interfaces: Learner, Probabilistic, Verbose, Persistable
Compatibility: Categorical, Continuous

Parameters:

# Param Default Type Description
1 max depth PHP_INT_MAX int The maximum depth of a branch.
2 max leaf size 3 int The max number of samples that a leaf node can contain.
3 min purity increase 0. float The minimum increase in purity necessary for a node not to be post pruned.
4 max features Auto int The max number of features to consider when determining a best split.
5 tolerance 1e-3 float A small amount of impurity to tolerate when choosing a best split.

Additional Methods:

Return the feature importances calculated during training indexed by feature column:

public featureImportances() : array

Return the height of the tree:

public height() : int

Return the balance of the tree:

public balance() : int

Example:

use Rubix\ML\Classifiers\ExtraTreeClassifier;

$estimator = new ExtraTreeClassifier(50, 3, 0.10, 4, 1e-3);

References:

  • P. Geurts et al. (2005). Extremely Randomized Trees.

Gaussian Naive Bayes

A variate of the Naive Bayes algorithm that uses a probability density function (PDF) over continuous features that are assumed to be normally distributed.

Interfaces: Learner, Online, Probabilistic, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 priors Auto array The user-defined class prior probabilities as an associative array with class labels as keys and the prior probabilities as values.

Additional Methods:

Return the class prior probabilities:

public priors() : array

Return the running mean of each feature column for each class:

public means() : ?array

Return the running variance of each feature column for each class:

public variances() : ?array

Example:

use Rubix\ML\Classifiers\GaussianNB;

$estimator = new GaussianNB([
	'benign' => 0.9,
	'malignant' => 0.1,
]);

References:

  • T. F. Chan et al. (1979). Updating Formulae and a Pairwise Algorithm for Computing Sample Variances.

K-d Neighbors

A fast K Nearest Neighbors algorithm that uses a K-d tree to divide the training set into neighborhoods whose max size are controlled by the max leaf size parameter. K-d Neighbors does a binary search to locate the nearest neighborhood and then prunes all neighborhoods whose bounding box is further than the kth nearest neighbor found so far. The main advantage of K-d Neighbors over regular brute force KNN is that it is faster, however it cannot be partially trained.

Interfaces: Learner, Probabilistic, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 k 3 int The number of neighboring training samples to consider when making a prediction.
2 kernel Euclidean object The distance kernel used to compute the distance between sample points.
3 weighted true bool Should we use the inverse distances as confidence scores when making predictions?
4 max leaf size 30 int The max number of samples in a leaf node (neighborhood).

Additional Methods:

Return the height of the tree:

public height() : int

Return the balance of the tree:

public balance() : int

Example:

use Rubix\ML\Classifiers\KDNeighbors;
use Rubix\ML\Kernels\Distance\Euclidean;

$estimator = new KDNeighbors(3, new Euclidean(), false, 10);

K Nearest Neighbors

A distance-based algorithm that locates the K nearest neighbors from the training set and uses a weighted vote to classify the unknown sample.

Note: K Nearest Neighbors is considered a lazy learner because it does the majority of its computation at inference. For a fast tree-based version, see KD Neighbors.

Interfaces: Learner, Online, Probabilistic, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 k 3 int The number of neighboring training samples to consider when making a prediction.
2 kernel Euclidean object The distance kernel used to compute the distance between sample points.
3 weighted true bool Should we use the inverse distances as confidence scores when making predictions?

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\Classifiers\KNearestNeighbors;
use Rubix\ML\Kernels\Distance\Manhattan;

$estimator = new KNearestNeighbors(3, new Manhattan(), true);

Logistic Regression

A type of linear classifier that uses the logistic (sigmoid) function to estimate the probabilities of exactly two classes.

Interfaces: Learner, Online, Probabilistic, Verbose, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 batch size 100 int The number of training samples to process at a time.
2 optimizer Adam object The gradient descent optimizer used to train the underlying network.
3 alpha 1e-4 float The amount of L2 regularization to apply to the weights of the network.
4 epochs 1000 int The maximum number of training epochs to execute.
5 min change 1e-4 float The minimum change in the cost function necessary to continue training.
6 cost fn Cross Entropy object The function that computes the cost of an erroneous activation during training.

Additional Methods:

Return the average loss of a sample at each epoch of training:

public steps() : array

Return the underlying neural network instance or null if untrained:

public network() : Network|null

Example:

use Rubix\ML\Classifers\LogisticRegression;
use Rubix\ML\NeuralNet\Optimizers\Adam;
use Rubix\ML\NeuralNet\CostFunctions\CrossEntropy;

$estimator = new LogisticRegression(10, new Adam(0.001), 1e-4, 100, 1e-4, new CrossEntropy());

Multi Layer Perceptron

A multiclass feedforward Neural Network classifier that uses a series of user-defined hidden layers as intermediate computational units. Multiple layers and non-linear activation functions allow the Multi Layer Perceptron to handle complex deep learning problems.

Note: The MLP features progress monitoring which stops training early if it can no longer make progress. It also utilizes snapshotting to make sure that it always has the best settings of the model parameters even if progress began to decline during training.

Interfaces: Learner, Online, Probabilistic, Verbose, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 hidden array An array composing the hidden layers of the neural network.
2 batch size 100 int The number of training samples to process at a time.
3 optimizer Adam object The gradient descent optimizer used to train the underlying network.
4 alpha 1e-4 float The amount of L2 regularization to apply to the weights of the network.
5 epochs 1000 int The maximum number of training epochs to execute.
6 min change 1e-4 float The minimum change in the cost function necessary to continue training.
7 cost fn Cross Entropy object The function that computes the cost of an erroneous activation during training.
8 holdout 0.1 float The ratio of samples to hold out for progress monitoring.
9 metric F1 Score object The validation metric used to monitor the training progress of the network.
10 window 3 int The number of epochs to consider when determining if the algorithm should terminate or keep training.

Additional Methods:

Return the average loss of a sample at each epoch of training:

public steps() : array

Return the validation scores at each epoch of training:

public scores() : array

Returns the underlying neural network instance or null if untrained:

public network() : Network|null

Example:

use Rubix\ML\Classifiers\MultiLayerPerceptron;
use Rubix\ML\NeuralNet\Layers\Dense;
use Rubix\ML\NeuralNet\Layers\Dropout;
use Rubix\ML\NeuralNet\Layers\Activation;
use Rubix\ML\NeuralNet\ActivationFunctions\LeakyReLU;
use Rubix\ML\NeuralNet\ActivationFunctions\PReLU;
use Rubix\ML\NeuralNet\Optimizers\Adam;
use Rubix\ML\NeuralNet\CostFunctions\CrossEntropy;
use Rubix\ML\CrossValidation\Metrics\MCC;

$estimator = new MultiLayerPerceptron([
	new Dense(30),
	new Activation(new LeakyReLU()),
	new Dropout(0.3),
	new Dense(20),
	new Activation(new LeakyReLU()),
	new Dropout(0.2),
	new Dense(10),
	new PReLU(0.25),
], 100, new Adam(0.001), 1e-4, 1000, 1e-3, new CrossEntropy(), 0.1, new MCC(), 3);

References:

  • G. E. Hinton. (1989). Connectionist learning procedures.

Naive Bayes

Probability-based classifier that estimates posterior probabilities of each class using Bayes' Theorem and the conditional probabilities calculated during training. The naive part relates to the fact that the algorithm assumes that all features are independent (non-correlated).

Interfaces: Learner, Online, Probabilistic, Persistable
Compatibility: Categorical

Parameters:

# Param Default Type Description
1 alpha 1.0 float The amount of additive (Laplace/Lidstone) smoothing to apply to the probabilities.
2 priors Auto array The class prior probabilities as an associative array with class labels as keys and the prior probabilities as values.

Additional Methods:

Return the class prior probabilities:

public priors() : array

Return the negative log probabilities of each feature given each class label:

public probabilities() : array

Example:

use Rubix\ML\Classifiers\NaiveBayes;

$estimator = new NaiveBayes(2.5, [
	'spam' => 0.3,
	'not spam' => 0.7,
]);

Radius Neighbors

Radius Neighbors is a spatial tree-based classifier that takes the weighted vote of each neighbor within a fixed user-defined radius measured by a kernel distance function.

Note: Unknown samples with 0 samples from the training set that are within radius will be labeled as outliers (-1).

Interfaces: Learner, Probabilistic, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 radius 1.0 float The radius within which points are considered neighboors.
2 kernel Euclidean object The distance kernel used to compute the distance between sample points.
3 weighted true bool Should we use the inverse distances as confidence scores when making predictions?
4 max leaf size 30 int The max number of samples in a leaf node (ball).

Additional Methods:

Return the height of the tree:

public height() : int

Return the balance of the tree:

public balance() : int

Example:

use Rubix\ML\Classifiers\RadiusNeighbors;
use Rubix\ML\Kernels\Distance\Manhattan;

$estimator = new RadiusNeighbors(50.0, new Manhattan(), false, 30);

Random Forest

Ensemble classifier that trains Decision Trees (Classification Trees or Extra Trees) on a random subset (bootstrap set) of the training data. A prediction is made based on the probability scores returned from each tree in the forest averaged and weighted equally.

Interfaces: Learner, Probabilistic, Persistable
Compatibility: Categorical, Continuous

Parameters:

# Param Default Type Description
1 base Classification Tree object The base tree estimator.
2 estimators 100 int The number of estimators to train in the ensemble.
3 ratio 0.1 float The ratio of random samples to train each estimator with.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\Classifiers\RandomForest;
use Rubix\ML\Classifiers\ClassificationTree;

$estimator = new RandomForest(ClassificationTree(10), 400, 0.1);

References:

  • L. Breiman. (2001). Random Forests.
  • L. Breiman et al. (2005). Extremely Randomized Trees.

Softmax Classifier

A generalization of Logistic Regression for multiclass problems using a single layer neural network with a Softmax output layer.

Interfaces: Learner, Online, Probabilistic, Verbose, Persistable
Compatibility: Continous

Parameters:

# Param Default Type Description
1 batch size 100 int The number of training samples to process at a time.
2 optimizer Adam object The gradient descent optimizer used to train the underlying network.
3 alpha 1e-4 float The amount of L2 regularization to apply to the weights of the network.
4 epochs 1000 int The maximum number of training epochs to execute.
5 min change 1e-4 float The minimum change in the cost function necessary to continue training.
6 cost fn Cross Entropy object The function that computes the cost of an erroneous activation during training.

Additional Methods:

Return the average loss of a sample at each epoch of training:

public steps() : array

Return the underlying neural network instance or null if untrained:

public network() : Network|null

Example:

use Rubix\ML\Classifiers\SoftmaxClassifier;
use Rubix\ML\NeuralNet\Optimizers\Momentum;
use Rubix\ML\NeuralNet\CostFunctions\CrossEntropy;

$estimator = new SoftmaxClassifier(256, new Momentum(0.001), 1e-4, 300, 1e-4, new CrossEntropy());

SVC

The multiclass Support Vector Machine (SVM) Classifier is a maximum margin classifier that can efficiently perform non-linear classification by implicitly mapping feature vectors into high dimensional feature space (called the kernel trick).

Note: This estimator requires the SVM PHP extension which uses the LIBSVM engine written in C++ under the hood.

Interfaces: Learner, Persistable
Compatibility: Continous

Parameters:

# Param Default Type Description
1 c 1.0 float The parameter that defines the width of the margin used to separate the classes.
2 kernel RBF object The kernel function used to operate in higher dimensions.
3 shrinking true bool Should we use the shrinking heuristic?
4 tolerance 1e-3 float The minimum change in the cost function necessary to continue training.
5 cache size 100. float The size of the kernel cache in MB.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\Classifiers\SVC;
use Rubix\ML\Kernels\SVM\Linear;

$estimator = new SVC(1.0, new Linear(), true, 1e-3, 100.);

References:

  • C. Chang et al. (2011). LIBSVM: A library for support vector machines.

Clusterers

Clustering is a technique in machine learning that focuses on grouping samples in such a way that the groups are similar. Another way of looking at it is that clusterers take unlabeled data points and assign them a label (cluster number).

DBSCAN

Density-Based Spatial Clustering of Applications with Noise is a clustering algorithm able to find non-linearly separable and arbitrarily-shaped clusters. In addition, DBSCAN also has the ability to mark outliers as noise and thus can be used as a quasi Anomaly Detector.

Note: Noise samples are assigned the cluster number -1.

Interfaces: None
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 radius 0.5 float The maximum radius between two points for them to be considered in the same cluster.
2 min density 5 int The minimum number of points within radius of each other to form a cluster.
3 kernel Euclidean object The distance kernel used to compute the distance between sample points.
4 max leaf size 30 int The max number of samples in a leaf node (ball).

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\Clusterers\DBSCAN;
use Rubix\ML\Kernels\Distance\Diagonal;

$estimator = new DBSCAN(4.0, 5, new Diagonal(), 20);

References:

  • M. Ester et al. (1996). A Densty-Based Algorithm for Discovering Clusters.

Fuzzy C Means

Distance-based soft clusterer that allows samples to belong to multiple clusters if they fall within a fuzzy region controlled by the fuzz parameter.

Interfaces: Learner, Probabilistic, Verbose, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 c int The number of target clusters.
2 fuzz 2.0 float Determines the bandwidth of the fuzzy area.
3 kernel Euclidean object The distance kernel used to compute the distance between sample points.
4 epochs 300 int The maximum number of training rounds to execute.
5 min change 1e-4 float The minimum change in inter cluster distance necessary for the algorithm to continue training.
6 seeder PlusPlus object The seeder used to initialize the cluster centroids.

Additional Methods:

Return the c computed centroids of the training set:

public centroids() : array

Returns the inter-cluster distances at each epoch of training:

public steps() : array

Example:

use Rubix\ML\Clusterers\FuzzyCMeans;
use Rubix\ML\Kernels\Distance\Euclidean;
use Rubix\ML\Clusterers\Seeders\Random;

$estimator = new FuzzyCMeans(5, 1.2, new Euclidean(), 300, 1e-3, new Random());

References:

  • J. C. Bezdek et al. (1984). FCM: The Fuzzy C-Means Clustering Algorithm.

Gaussian Mixture

A Gaussian Mixture model (GMM) is a probabilistic model for representing the presence of clusters within an overall population without requiring a sample to know which sub-population it belongs to a priori. GMMs are similar to centroid-based clusterers like K Means but allow both the centers (means) and the radii (variances) to be learned as well.

Interfaces: Learner, Probabilistic, Verbose, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 k int The number of target clusters.
2 min change 1e-3 float The minimum change in the Gaussians necessary for the algorithm to continue training.
3 epochs 100 int The maximum number of training rounds to execute.

Additional Methods:

Return the cluster prior probabilities based on their representation over all training samples:

public priors() : array

Return the running means of each feature column for each cluster:

public means() : array

Return the variance of each feature column for each cluster:

public variances() : array

Example:

use Rubix\ML\Clusterers\FuzzyCMeans;
use Rubix\ML\Kernels\Distance\Euclidean;
use Rubix\ML\Clusterers\Seeders\PlusPlus;

$estimator = new FuzzyCMeans(5, 1.2, new Euclidean(), 1e-3, 1000, new PlusPlus());

References:

  • A. P. Dempster et al. (1977). Maximum Likelihood from Incomplete Data via the EM Algorithm.
  • J. Blomer et al. (2016). Simple Methods for Initializing the EM Algorithm for Gaussian Mixture Models.

K Means

A fast online centroid-based hard clustering algorithm capable of clustering linearly separable data points given some prior knowledge of the target number of clusters (defined by k). K Means is trained with mini batch gradient descent using the within cluster distance as a loss function.

Interfaces: Learner, Persistable, Verbose
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 k int The number of target clusters.
2 batch size 100 int The size of each mini batch in samples.
3 kernel Euclidean object The distance kernel used to compute the distance between sample points.
4 epochs 300 int The maximum number of training rounds to execute.
5 min change 1 int The minimum change in the size of each cluster for training to continue.
6 seeder PlusPlus object The seeder used to initialize the cluster centroids.

Additional Methods:

Return the k computed centroids of the training set:

public centroids() : array

Example:

use Rubix\ML\Clusterers\KMeans;
use Rubix\ML\Kernels\Distance\Euclidean;

$estimator = new KMeans(3, 100, new Euclidean(), 300, 1);

References:

  • D. Sculley. (2010). Web-Scale K-Means Clustering.

Mean Shift

A hierarchical clustering algorithm that uses peak finding to locate the local maxima (centroids) of a training set given by a radius constraint.

Interfaces: Learner, Verbose, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 radius float The radius of each cluster centroid.
2 kernel Euclidean object The distance kernel used to compute the distance between samples.
3 threshold 1e-8 float The minimum change in centroid means necessary for the algorithm to continue training.
4 epochs 100 int The maximum number of training rounds to execute.

Additional Methods:

Return the centroids computed from the training set:

public centroids() : array

Returns the amount of centroid shift during each epoch of training:

public steps() : array

Example:

use Rubix\ML\Clusterers\MeanShift;
use Rubix\ML\Kernels\Distance\Diagonal;

$estimator = new MeanShift(3.0, new Diagonal(), 1e-6, 2000);

References:

  • M. A. Carreira-Perpinan et al. (2015). A Review of Mean-shift Algorithms for Clustering.

Seeders

Seeders are responsible for initializing the cluster centroids of certain learners.

Plus Plus

This seeder attempts to maximize the likelihood of seeding distant clusters while still remaining random. It does so by sequentially selecting random samples weighted by their distance from the previous seed.

Parameters:

# Param Default Type Description
1 kernel Euclidean object The distance kernel used to compute the distance between samples.

Example:

use Rubix\ML\Clusterers\Seeders\PlusPlus;
use Rubix\ML\Kernels\Distance\Minkowski;

$seeder = new PlusPlus(new Minkowski(5.0));

References:

  • D. Arthur et al. (2006). k-means++: The Advantages of Careful Seeding.
  • A. Stetco et al. (2015). Fuzzy C-means++: Fuzzy C-means with effective seeding initialization.

Random

Completely random selection of k seeds from the given dataset.

Parameters:

This seeder does not have any parameters.

Example:

use Rubix\ML\Clusterers\Seeders\Random;

$seeder = new Random();

Embedders

Manifold learning is a type of non-linear dimensionality reduction used primarily for visualizing high dimensional datasets in low (1 to 3) dimensions. Embedders are manifold learners that provide the predict() API for embedding a dataset. The predictions of an Embedder are the low dimensional embeddings as n-dimensional arrays where n is the dimensionality of the embedding.

t-SNE

T-distributed Stochastic Neighbor Embedding is a two-stage non-linear manifold learning algorithm based on batch Gradient Descent. During the first stage (early stage) the samples are exaggerated to encourage distant clusters. Since the t-SNE cost function (KL Divergence) has a rough gradient, momentum is employed to help escape bad local minima.

Interfaces: Verbose
Compatibility: Continous

Parameters:

# Param Default Type Description
1 dimensions 2 int The number of dimensions of the target embedding.
2 perplexity 30 int The number of effective nearest neighbors to refer to when computing the variance of the Gaussian over that sample.
3 exaggeration 12. float The factor to exaggerate the distances between samples during the early stage of fitting.
4 rate 100. float The learning rate that controls the step size.
5 kernel Euclidean object The distance kernel to use when measuring distances between samples.
6 epochs 1000 int The number of times to iterate over the embedding.
7 min gradient 1e-8 float The minimum gradient necessary to continue embedding.
8 window 3 int The number of most recent epochs to consider when determining an early stop.

Additional Methods:

Return the magnitudes of the gradient at each epoch from the last embedding:

public steps() : array

Example:

use Rubi\ML\Embedders\TSNE;
use Rubix\ML\Kernels\Manhattan;

$embedder = new TSNE(2, 30, 12., 10., new Manhattan(), 500, 1e-6, 5);

References:

  • L. van der Maaten et al. (2008). Visualizing Data using t-SNE.
  • L. van der Maaten. (2009). Learning a Parametric Embedding by Preserving Local Structure.

Regressors

Regressors are used to predict continuous real-valued outcomes.

Adaline

Adaptive Linear Neuron or (Adaline) is a type of single layer neural network with a linear output neuron. Training is equivalent to solving Ridge regression iteratively using mini batch Gradient Descent.

Interfaces: Learner, Online, Verbose, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 batch size 100 int The number of training samples to process at a time.
2 optimizer Adam object The gradient descent optimizer used to train the underlying network.
3 alpha 1e-4 float The amount of L2 regularization to apply to the weights of the network.
4 epochs 100 int The maximum number of training epochs to execute.
5 min change 1e-4 float The minimum change in the cost function necessary to continue training.
6 cost fn Least Squares object The function that computes the cost of an erroneous activation during training.

Additional Methods:

Return the average loss of a sample at each epoch of training:

public steps() : array

Return the underlying neural network instance or null if untrained:

public network() : Network|null

Example:

use Rubix\ML\Classifers\Adaline;
use Rubix\ML\NeuralNet\Optimizers\Adam;
use Rubix\ML\NeuralNet\CostFunctions\HuberLoss;

$estimator = new Adaline(10, new Adam(0.001), 500, 1e-6, new HuberLoss(2.5));

Dummy Regressor

Regressor that guesses output values based on a user-defined Guessing Strategy. Dummy Regressor is useful to provide a sanity check and to compare performance against actual Regressors.

Interfaces: Learner, Persistable
Compatibility: Categorical, Continuous, Resource

Parameters:

# Param Default Type Description
1 strategy Mean object The guessing strategy to employ when guessing the outcome of a sample.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\Regressors\DummyRegressor;
use Rubix\ML\Other\Strategies\BlurryPercentile;

$estimator = new DummyRegressor(new BlurryPercentile(56.5, 0.1));

Extra Tree Regressor

An Extremely Randomized Regression Tree, these trees differ from standard Regression Trees in that they choose a split drawn from a random set determined by the max features parameter, rather than searching the entire column for the best split.

Note: Decision tree based algorithms can handle both categorical and continuous features at the same time.

Interfaces: Learner, Verbose, Persistable
Compatibility: Categorical, Continuous

Parameters:

# Param Default Type Description
1 max depth PHP_INT_MAX int The maximum depth of a branch that is allowed.
2 max leaf size 3 int The max number of samples that a leaf node can contain.
3 min purity increase 0. float The minimum increase in purity necessary for a node not to be post pruned.
4 max features Auto int The number of features to consider when determining a best split.
5 tolerance 1e-4 float A small amount of impurity to tolerate when choosing a best split.

Additional Methods:

Return the feature importances calculated during training indexed by feature column:

public featureImportances() : array

Return the height of the tree:

public height() : int

Return the balance of the tree:

public balance() : int

References:

  • P. Geurts et al. (2005). Extremely Randomized Trees.

Example:

use Rubix\ML\Classifiers\ExtraTreeRegressor;

$estimator = new ExtraTreeRegressor(30, 3, 0.05, 20, 1e-4);

Gradient Boost

Gradient Boost is a stage-wise additive model that uses a Gradient Descent boosting paradigm for training boosters (Regression Trees) to correct the error residuals of a weak base learner.

Note: The default base regressor is a Dummy Regressor using the Mean Strategy and the default booster is a Regression Tree with a max depth of 3.

Interfaces: Learner, Ensemble, Verbose, Persistable
Compatibility: Depends on base learner

Parameters:

# Param Default Type Description
1 booster Regression Tree object The regressor that will fix up the error residuals of the base learner.
2 rate 0.1 float The learning rate of the ensemble.
3 estimators 100 int The number of estimators to train in the ensemble.
4 ratio 0.8 float The ratio of samples to subsample from the training dataset per epoch.
5 min change 1e-4 float The minimum change in the cost function necessary to continue training.
6 tolerance 1e-3 float The amount of mean squared error to tolerate before early stopping.
7 base Dummy Regressor object The weak learner to be boosted.

Additional Methods:

Return the training error at each epoch:

public steps() : array

Example:

use Rubix\ML\Regressors\GradientBoost;
use Rubix\ML\Regressors\DummyRegressor;
use Rubix\ML\Regressors\RegressionTree;
use Rubix\ML\Other\Strategies\Mean;

$estimator = new GradientBoost(new RegressionTree(3), 0.1, 400, 0.3, 1e-4, 1e-3, new DummyRegressor(new Mean()));

References:

  • J. H. Friedman. (2001). Greedy Function Approximation: A Gradient Boosting Machine.

K-d Neighbors Regressor

A fast implementation of KNN Regressor using a spatially-aware K-d tree. The KDN Regressor works by locating the neighborhood of a sample via binary search and then does a brute force search only on the samples close to or within the neighborhood. The main advantage of K-d Neighbors over brute force KNN is inference speed, however you no longer have the ability to partially train.

Interfaces: Learner, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 k 3 int The number of neighboring training samples to consider when making a prediction.
2 kernel Euclidean object The distance kernel used to compute the distance between sample points.
3 weighted true bool Should we use the inverse distances as confidence scores when making predictions?
4 max leaf size 30 int The max number of samples in a leaf node (neighborhood).

Additional Methods:

Return the height of the tree:

public height() : int

Return the balance of the tree:

public balance() : int

Example:

use Rubix\ML\Regressors\KDNeighborsRegressor;
use Rubix\ML\Kernels\Distance\Minkowski;

$estimator = new KDNeighborsRegressor(5, new Minkowski(4.0), true, 30);

KNN Regressor

A version of K Nearest Neighbors that uses the average (mean) outcome of K nearest data points to make continuous valued predictions suitable for regression problems.

Note: K Nearest Neighbors is considered a lazy learning estimator because it does the majority of its computation at prediction time.

Interfaces: Learner, Online, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 k 3 int The number of neighboring training samples to consider when making a prediction.
2 kernel Euclidean object The distance kernel used to compute the distance between sample points.
3 weighted true bool Should we use the inverse distances as confidence scores when making predictions?

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\Regressors\KNNRegressor;
use Rubix\ML\Kernels\Distance\Minkowski;

$estimator = new KNNRegressor(2, new Minkowski(3.0), false);

MLP Regressor

A multi layer feedforward Neural Network with a continuous output layer suitable for regression problems. Like the Multi Layer Perceptron classifier, the MLP Regressor is able to tackle deep learning problems by forming higher-order representations of the features using intermediate computational units called hidden layers.

Note: The MLP features progress monitoring which stops training early if it can no longer make progress. It also utilizes snapshotting to make sure that it always has the best settings of the model parameters even if progress began to decline during training.

Interfaces: Learner, Online, Verbose, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 hidden array An array composing the hidden layers of the neural network.
2 batch size 100 int The number of training samples to process at a time.
3 optimizer Adam object The gradient descent optimizer used to train the underlying network.
4 alpha 1e-4 float The amount of L2 regularization to apply to the weights of the network.
5 epochs 1000 int The maximum number of training epochs to execute.
6 min change 1e-4 float The minimum change in the cost function necessary to continue training.
7 cost fn Least Squares object The function that computes the cost of an erroneous activation during training.
8 holdout 0.1 float The ratio of samples to hold out for progress monitoring.
9 metric Mean Squared Error object The validation metric used to monitor the training progress of the network.
10 window 3 int The number of epochs to consider when determining if the algorithm should terminate or keep training.

Additional Methods:

Return the average loss of a sample at each epoch of training:

public steps() : array

Return the validation scores at each epoch of training:

public scores() : array

Returns the underlying neural network instance or null if untrained:

public network() : Network|null

Example:

use Rubix\ML\Regressors\MLPRegressor;
use Rubix\ML\NeuralNet\Layers\Dense;
use Rubix\ML\NeuralNet\Layers\Activation;
use Rubix\ML\NeuralNet\ActivationFunctions\LeakyReLU;
use Rubix\ML\NeuralNet\Optimizers\RMSProp;
use Rubix\ML\CrossValidation\Metrics\RSquared;

$estimator = new MLPRegressor([
	new Dense(50),
	new Activation(new LeakyReLU(0.1)),
	new Dense(50),
	new Activation(new LeakyReLU(0.1)),
	new Dense(50),
	new Activation(new LeakyReLU(0.1)),
], 256, new RMSProp(0.001), 1e-3, 100, 1e-5, new LeastSquares(), 0.1, new RSquared(), 3);

References:

  • G. E. Hinton. (1989). Connectionist learning procedures.

Radius Neighbors Regressor

This is the regressor version of Radius Neighbors classifier implementing a binary spatial tree under the hood for fast radius queries. The prediction is a weighted average of each label from the training set that is within a fixed user-defined radius.

Note: Unknown samples with 0 samples from the training set that are within radius will be labeled NaN.

Interfaces: Learner, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 radius 1.0 float The radius within which points are considered neighboors.
2 kernel Euclidean object The distance kernel used to compute the distance between sample points.
3 weighted true bool Should we use the inverse distances as confidence scores when making predictions?
4 max leaf size 30 int The max number of samples in a leaf node (ball).

Additional Methods:

Return the height of the tree:

public height() : int

Return the balance of the tree:

public balance() : int

Example:

use Rubix\ML\Regressors\RadiusNeighborsRegressor;
use Rubix\ML\Kernels\Distance\Diagonal;

$estimator = new RadiusNeighborsRegressor(0.5, new Diagonal(), true, 20);

Regression Tree

A Decision Tree learning algorithm (CART) that performs greedy splitting by minimizing the impurity (variance) of the labels at each decision node split.

Note: Decision tree based algorithms can handle both categorical and continuous features at the same time.

Interfaces: Learner, Verbose, Persistable
Compatibility: Categorical, Continuous

Parameters:

# Param Default Type Description
1 max depth PHP_INT_MAX int The maximum depth of a branch.
2 max leaf size 3 int The maximum number of samples that a leaf node can contain.
3 min purity increase 0. float The minimum increase in purity necessary for a node not to be post pruned.
4 max features Auto int The maximum number of features to consider when determining a best split.
5 tolerance 1e-4 float A small amount of impurity to tolerate when choosing a best split.

Additional Methods:

Return the feature importances calculated during training indexed by feature column:

public featureImportances() : array

Return the height of the tree:

public height() : int

Return the balance of the tree:

public balance() : int

Example:

use Rubix\ML\Regressors\RegressionTree;

$estimator = new RegressionTree(30, 2, 35., null, 1e-4);

Ridge

L2 penalized least squares linear regression solved using closed-form equation.

Interfaces: Learner, Persistable
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 alpha 1.0 float The L2 regularization penalty amount to be added to the weight coefficients.

Additional Methods:

Return the weights of the model:

public weights() : array|null

Return the bias parameter:

public bias() : float|null

Example:

use Rubix\ML\Regressors\Ridge;

$estimator = new Ridge(2.0);

SVR

The Support Vector Machine Regressor is a maximum margin algorithm for the purposes of regression. Similarly to the Support Vector Machine Classifier, the model produced by SVR (R for regression) depends only on a subset of the training data, because the cost function for building the model ignores any training data close to the model prediction given by parameter epsilon. Thus, the value of epsilon defines a margin of tolerance where no penalty is given to errors.

Note: This estimator requires the SVM PHP extension which uses the LIBSVM engine written in C++ under the hood.

Interfaces: Learner, Persistable
Compatibility: Continous

Parameters:

# Param Default Type Description
1 c 1.0 float The parameter that defines the width of the margin used to separate the classes.
2 epsilon 0.1 float Specifies the margin within which no penalty is associated in the training loss.
3 kernel RBF object The kernel function used to operate in higher dimensions.
4 shrinking true bool Should we use the shrinking heuristic?
5 tolerance 1e-3 float The minimum change in the cost function necessary to continue training.
6 cache size 100. float The size of the kernel cache in MB.

Additional Methods:

This estimator does not have any additional methods.

Example:

use Rubix\ML\Classifiers\SVC;
use Rubix\ML\Kernels\SVM\Linear;

$estimator = new SVR(1.0, 0.03, new RBF(), true, 1e-3, 256.);

Refernces:

  • C. Chang et al. (2011). LIBSVM: A library for support vector machines.
  • A. Smola et al. (2003). A Tutorial on Support Vector Regression.

Meta-Estimators

Meta-estimators enhance base estimators by adding additional functionality such as data preprocessing, model persistence, and model averaging. Meta-estimators take on the type (Classifier, Regressor, etc.) of the base estimator they wrap and allow methods on the base estimator to be called from the parent.

Bootstrap Aggregator

Bootstrap Aggregating (or bagging for short) is a model averaging technique designed to improve the stability and performance of a user-specified base estimator by training a number of them on a unique bootstrapped training set sampled at random with replacement.

Note: Bootstrap Aggregator does not work with clusterers or embedders.

Interfaces: Learner | Persistable
Compatibility: Depends on base learner

Parameters:

# Param Default Type Description
1 base object The base estimator to be used in the ensemble.
2 estimators 10 int The number of base estimators to train in the ensemble.
3 ratio 0.5 float The ratio of samples from the training set to train each base estimator with.

Additional Methods:

This meta estimator does not have any additional methods.

Example:

use Rubix\ML\BootstrapAggregator;
use Rubix\ML\Regressors\RegressionTree;

$estimator = new BootstrapAggregator(new RegressionTree(5), 100, 0.2);

References:

  • L. Breiman. (1996). Bagging Predictors.

Grid Search

Grid Search is an algorithm that optimizes hyper-parameter selection. From the user's perspective, the process of training and predicting is the same, however, under the hood, Grid Search trains one estimator per combination of parameters and the best model is selected as the base estimator.

Note: You can choose the parameters to search manually or you can generate them randomly or in a grid using the Params helper.

Interfaces: Learner, Persistable, Verbose
Compatibility: Depends on base learner

Parameters:

# Param Default Type Description
1 base string The fully qualified class name of the base Estimator.
2 grid array An array of n-tuples where each tuple contains possible parameters for a given constructor location by ordinal.
3 metric Auto object The validation metric used to score each set of hyper-parameters.
4 validator KFold object An instance of a validator object (HoldOut, KFold, etc.) that will be used to test each model.
5 retrain true bool Should we retrain using the best parameter combination and entire dataset?

Additional Methods:

Return every parameter combination from the last grid search:

public params() : array

The validation scores of the last search:

public scores() : array

A tuple containing the best parameters and their validation score:

public best() : array

Return the underlying base estimator:

public estimator() : Estimator

Example:

use Rubix\ML\GridSearch;
use Rubix\ML\Classifiers\KNearestNeighbors;
use Rubix\ML\Kernels\Distance\Euclidean;
use Rubix\ML\Kernels\Distance\Manhattan;
use Rubix\ML\CrossValidation\Metrics\F1Score;
use Rubix\ML\CrossValidation\KFold;

$grid = [
	[1, 3, 5, 10], [new Euclidean(), new Manhattan()], [true, false],
];

$estimator = new GridSearch(KNearestNeightbors::class, $grid, new F1Score(), new KFold(10), true);

Model Orchestra

A Model Orchestra is a stacked model ensemble comprised of an orchestra of estimators (Classifiers or Regressors) and a conductor estimator. The role of the conductor is to learn the influence scores of each estimator in the orchestra while using their predictions as inputs to make a final weighted prediction.

Note: The features that each estimator passes on to the conductor may vary depending on the type of estimator. For example, a Probabilistic classifier will pass class probability scores while a regressor will pass on a single real value. If a datatype is not compatible with the conducting estimator, then wrap it in a Pipeline and use a transformer such as One Hot Encoder or Interval Discretizer.

Interfaces: Learner, Probabilistic, Persistable, Verbose
Compatibility: Depends on base learners

Parameters:

# Param Default Type Description
1 orchestra array The estimator instances that comprise the orchestra section of the ensemble.
2 conductor object The estimator that will weight each prediction and give the final output.
3 ratio 0.8 float The ratio of samples used to train the orchestra (the remaining are used to train the conductor).

Additional Methods:

Return an array of estimators comprising the orchestra part of the ensemble:

public orchestra() : array

Return the conductor estimator:

public conductor() : Estimator

Example:

use Rubix\ML\ModelOrchestra;
use Rubix\ML\Classifiers\GaussianNB;
use Rubix\ML\Classifiers\KNearestNeighbors;
use Rubix\ML\Classifiers\ClassificationTree;
use Rubix\ML\Classifiers\SoftmaxClassifier;

$estimator = new ModelOrchestra([
	new ClassificationTree(10, 3, 2),
	new KNearestNeighbors(3, new Euclidean()),
	new GaussianNB(),
], new SoftmaxClassifier(10), 0.8);

Persistent Model

It is possible to persist a model by wrapping the estimator instance in a Persistent Model meta-estimator. The Persistent Model wrapper gives the estimator three additional methods save(), load(), and prompt() that allow the estimator to be saved and retrieved from storage.

Interfaces: Learner, Probabilistic, Verbose
Compatibility: Depends on base learner

Parameters:

# Param Default Type Description
1 base object An instance of the base estimator to be persisted.
2 persister object The persister object used to store the model data.

Additional Methods:

Save the persistent model to storage:

public save() : void

Load the persistent model from storage given a persister:

public static load(Persister $persister) : self

Prompt the user to save the model or not via stdout:

public prompt() : void

Example:

use Rubix\ML\PersistentModel;
use Rubix\ML\Classifiers\LogisticRegression;
use Rubix\ML\NeuralNet\Optimizers\Adam;
use Rubix\ML\Persisters\Filesystem;
use Rubix\ML\Persisters\Serializers\Native;

$persister = new Filesystem('/random_forest.model', 2, new Native());

$estimator = new PersistentModel(new LogisticRegression(256, new Adam(0.001)), $persister);

Pipeline

Pipeline is a meta estimator responsible for transforming the input data by applying a series of transformer middleware. Pipeline accepts a base estimator and a list of transformers to apply to the input data before it is fed to the estimator. Under the hood, Pipeline will automatically fit the training set upon training and transform any Dataset object supplied as an argument to one of the base Estimator's methods, including train() and predict(). With the elastic mode enabled, Pipeline can update the fitting of certain transformers during online (partial) training.

Note: Since transformations are applied to dataset objects in place (without making a copy), using the dataset in a program after it has been run through Pipeline may have unexpected results. If you need a clean dataset object to call multiple methods with, you can use the PHP clone syntax to keep an original (untransformed) copy in memory.

Interfaces: Learner, Online, Persistable, Verbose
Compatibility: Depends on base learner and transformers

Parameters:

# Param Default Type Description
1 transformers array The transformer middleware to be applied to the input data in order.
2 estimator object An instance of the base estimator to receive transformed data.
3 elastic true bool Should we update the elastic transformers during partial training?

Additional Methods:

This meta estimator does not have any additional methods.

Example:

use Rubix\ML\Pipeline;
use Rubix\ML\Classifiers\SoftmaxClassifier;
use Rubix\ML\NeuralNet\Optimizer\Adam;
use Rubix\ML\Transformers\MissingDataImputer;
use Rubix\ML\Transformers\OneHotEncoder;
use Rubix\ML\Transformers\PrincipalComponentAnalysis;
use Rubix\ML\Transformers\ZScaleStandardizer;

$estimator = new Pipeline([
	new MissingDataImputer('?'),
	new OneHotEncoder(),
	new PrincipalComponentAnalysis(20),
	new ZScaleStandardizer(true),
], new SoftmaxClassifier(128, new Adam(0.001)), true);

Transformers

Transformers take Dataset objects and apply blanket transformations to the samples contained within them. They are often used as part of a Pipeline or they can be used by themselves. Examples of transformations are scaling, centering, normalization, dimensionality reduction, missing data imputation, and feature selection.

The transformer directly transforms the data in place via the transform() method:

public transform(array &$samples, ?array &$labels = null) : void

Note: To transform a dataset without having to pass the raw samples and labels you can pass a transformer to the apply() method on a Dataset object.

Stateful

For stateful transformers, the fit() method will allow the transformer to compute any necessary information from the training set in order to carry out its future transformations. You can think of fitting a transformer like training a learner.

To fit the transformer to a training set:

public fit(Dataset $dataset) : void

Check if the transformer has been fitted:

public fitted() : bool

Example

use Rubix\ML\Transformers\OneHotEncoder;

$transformer = new OneHotEncoder();

$transformer->fit($dataset);

Elastic

Some transformers are able to adapt to new training data. The update() method on transformers that implement the Elastic interface can be used to modify the fitting of the transformer with new data even after it has previously been fitted. Updating is to transformer as partially training is to online learner.

public update(Dataset $dataset) : void

Example

use Rubix\ML\Transformers\ZScaleStandardizer;

$transformer = new ZScaleStandardizer();

$folds = $dataset->fold(3);

$transformer->fit($folds[0]);

$transformer->update($folds[1]);

$transformer->update($folds[2]);

Dense Random Projector

The Dense Random Projector uses a random matrix sampled from a dense uniform distribution [-1, 1] to reduce the dimensionality of a dataset by projecting it onto a vector space of target dimensionality.

Interfaces: Stateful
Compatibility: Continuous only

Parameters:

# Param Default Type Description
1 dimensions int The number of target dimensions to project onto.

Additional Methods:

Estimate the minimum dimensionality needed given total sample size and max distortion using the Johnson-Lindenstrauss lemma:

public static estimate(int $n, float $maxDistortion = 0.1) : int

Example:

use Rubix\ML\Transformers\DenseRandomProjector;

$transformer = new DenseRandomProjector(50);

References:

  • D. Achlioptas. (2003). Database-friendly random projections: Johnson-Lindenstrauss with binary coins.

Gaussian Random Projector

A random projector is a dimensionality reducer based on the Johnson-Lindenstrauss lemma that uses a random matrix to project feature vectors onto a user-specified number of dimensions. It is faster than most non-randomized dimensionality reduction techniques such as PCA or LDA and it offers similar results. This version utilizes a random matrix sampled from a smooth Gaussian distribution.

Interfaces: Stateful
Compatibility: Continuous only

Parameters:

# Param Default Type Description
1 dimensions int The number of target dimensions to project onto.

Additional Methods:

Estimate the minimum dimensionality needed given total sample size and max distortion using the Johnson-Lindenstrauss lemma:

public static estimate(int $n, float $maxDistortion = 0.1) : int

Example:

use Rubix\ML\Transformers\GaussianRandomProjector;

$transformer = new GaussianRandomProjector(100);

HTML Stripper

Removes any HTML tags that may be in the text of a categorical variable.

Interfaces: None
Compatibility: Categorical

Parameters:

This transformer does not have any parameters.

Additional Methods:

This transformer does not have any additional methods.

Example:

use Rubix\ML\Transformers\HTMLStripper;

$transformer = new HTMLStripper();

Image Vectorizer

Image Vectorizer takes images (as PHP Resources) and converts them into a flat vector of raw color channel data. Scaling and cropping is handled automatically by Intervention Image for PHP.

Note: Note that the GD extension is required to use this transformer.

Interfaces: None
Compatibility: Resource (Images)

Parameters:

# Param Default Type Description
1 size [32, 32] array A tuple of width and height values denoting the resolution of the encoding.
2 rgb true bool True to use RGB color channel data and false to use greyscale.
3 driver 'gd' string The PHP extension to use for image processing ('gd' or 'imagick').

Additional Methods:

Return the dimensionality of the vector that gets encoded:

public dimensions() : int

Example:

use Rubix\ML\Transformers\ImageVectorizer;

$transformer = new ImageVectorizer([28, 28], true, 'gd');

Interval Discretizer

This transformer creates an equi-width histogram for each continuous feature column and encodes a discrete category with an automatic bin label. The Interval Discretizer is helpful when converting continuous features to categorical features so they can be learned by an estimator that supports categorical features natively.

Interfaces: Stateful
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 bins 5 int The number of bins (discrete features) per continuous feature column.

Additional Methods:

Return the possible categories of each feature column:

public categories() : array

Return the intervals of each continuous feature column calculated during fitting:

public intervals() : array

Example:

use Rubix\ML\Transformers\IntervalDiscretizer;

$transformer = new IntervalDiscretizer(10);

L1 Normalizer

Transform each sample vector in the sample matrix such that each feature is divided by the L1 norm (or magnitude) of that vector.

Interfaces: None
Compatibility: Continuous only

Parameters:

This transformer does not have any parameters.

Additional Methods:

This transformer does not have any additional methods.

Example:

use Rubix\ML\Transformers\L1Normalizer;

$transformer = new L1Normalizer();

L2 Normalizer

Transform each sample vector in the sample matrix such that each feature is divided by the L2 norm (or magnitude) of that vector.

Interfaces: None
Compatibility: Continuous only

Parameters:

This transformer does not have any parameters.

Additional Methods:

This transformer does not have any additional methods.

Example:

use Rubix\ML\Transformers\L2Normalizer;

$transformer = new L2Normalizer();

Lambda Function

Run a stateless lambda function (anonymous function) over the sample matrix. The lambda function receives the sample matrix (and labels if applicable) as an argument and should return the transformed sample matrix and labels in a 2-tuple.

Interfaces: None
Compatiblity: Depends on function

Parameters:

# Param Default Type Description
1 lambda callable The lambda function to run over the sample matrix.

Additional Methods:

This transformer does not have any additional methods.

Example:

use Rubix\ML\Transformers\LambdaFunction;

$transformer = new LambdaFunction(function ($samples, $labels) {
	$samples = array_map(function ($sample) {
		return [array_sum($sample)];
	}, $samples);

	return [$samples, $labels];
});

Linear Discriminant Analysis

A supervised dimensionality reduction technique that selects the most discriminating features based on class labels. In other words, LDA finds a linear combination of features that characterizes or best separates two or more classes.

Interfaces: Stateful
Compatibility: Continuous only

Parameters:

# Param Default Type Description
1 dimensions int The target number of dimensions to project onto.

Additional Methods:

Return the amount of variance that has been preserved by the transformation:

public explainedVar() : ?float

Return the amount of variance lost by discarding the noise components:

public noiseVar() : ?float

Return the percentage of information lost due to the transformation:

public lossiness() : ?float

Example:

use Rubix\ML\Transformers\LinearDiscriminantAnalysis;

$transformer = new LinearDiscriminantAnalysis(20);

Max Absolute Scaler

Scale the sample matrix by the maximum absolute value of each feature column independently such that the feature will be between -1 and 1.

Interfaces: Stateful, Elastic
Compatibility: Continuous

Parameters:

This transformer does not have any parameters.

Additional Methods:

Return the maximum absolute values for each feature column:

public maxabs() : array

Example:

use Rubix\ML\Transformers\MaxAbsoluteScaler;

$transformer = new MaxAbsoluteScaler();

Min Max Normalizer

The Min Max Normalizer scales the input features to a value between a user-specified range (default 0 to 1).

Interfaces: Stateful, Elastic
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 min 0. float The minimum value of the transformed features.
2 max 1. float The maximum value of the transformed features.

Additional Methods:

Return the minimum values for each fitted feature column:

public minimums() : ?array

Return the maximum values for each fitted feature column:

public maximums() : ?array

Example:

use Rubix\ML\Transformers\MinMaxNormalizer;

$transformer = new MinMaxNormalizer(-5., 5.);

Missing Data Imputer

In the real world, it is common to have data with missing values here and there. The Missing Data Imputer replaces missing value placeholder values with a guess based on a given Strategy.

Interfaces: Stateful
Compatibility: Categorical, Continuous

Parameters:

# Param Default Type Description
1 placeholder '?' string or numeric The placeholder value that denotes a missing value.
2 continuous strategy Mean object The guessing strategy to employ for continuous feature columns.
3 categorical strategy K Most Frequent object The guessing strategy to employ for categorical feature columns.

Additional Methods:

This transformer does not have any additional methods.

Example:

use Rubix\ML\Transformers\MissingDataImputer;
use Rubix\ML\Other\Strategies\BlurryPercentile;
use Rubix\ML\Other\Strategies\PopularityContest;

$transformer = new MissingDataImputer('?', new BlurryPercentile(0.61), new PopularityContest());

Numeric String Converter

Convert all numeric strings into their integer and floating point countertypes. Useful for when extracting from a source that only recognizes data as string types.

Interfaces: None
Compatibility: Categorical

Parameters:

This transformer does not have any parameters.

Additional Methods:

This transformer does not have any additional methods.

Example:

use Rubix\ML\Transformers\NumericStringConverter;

$transformer = new NumericStringConverter();

One Hot Encoder

The One Hot Encoder takes a column of categorical features and produces a n-d one-hot representation where n is equal to the number of unique categories in that column. A 0 in any location indicates that a category represented by that column is not present whereas a 1 indicates that a category is present in the sample.

Interfaces: Stateful
Compatibility: Categorical

Parameters:

This transformer does not have any parameters.

Additional Methods:

This transformer does not have any additional methods.

Example:

use Rubix\ML\Transformers\OneHotEncoder;

$transformer = new OneHotEncoder();

Polynomial Expander

This transformer will generate polynomials up to and including the specified degree of each continuous feature column. Polynomial expansion is sometimes used to fit data that is non-linear using a linear estimator such as Ridge or Logistic Regression.

Interfaces: None
Compatibility: Continuous only

Parameters:

# Param Default Type Description
1 degree 2 int The highest degree polynomial to generate from each feature vector.

Additional Methods:

This transformer does not have any additional methods.

Example:

use Rubix\ML\Transformers\PolynomialExpander;

$transformer = new PolynomialExpander(3);

Principal Component Analysis

Principal Component Analysis or PCA is a dimensionality reduction technique that aims to transform the feature space by the k principal components that explain the most variance of the data where k is the dimensionality of the output specified by the user. PCA is used to compress high dimensional samples down to lower dimensions such that they would retain as much of the information as possible.

Interfaces: Stateful
Compatibility: Continuous only

Parameters:

# Param Default Type Description
1 dimensions None int The target number of dimensions to project onto.

Additional Methods:

Return the amount of variance that has been preserved by the transformation:

public explainedVar() : ?float

Return the amount of variance lost by discarding the noise components:

public noiseVar() : ?float

Return the percentage of information lost due to the transformation:

public lossiness() : ?float

Example:

use Rubix\ML\Transformers\PrincipalComponentAnalysis;

$transformer = new PrincipalComponentAnalysis(15);

References:

  • H. Abdi et al. (2010). Principal Component Analysis.

Quartile Standardizer

This standardizer centers the dataset around its median and scales each feature according to the interquartile range (IQR) of that column. The IQR is defined as the range between the 1st quartile (25th quantile) and the 3rd quartile (75th quantile) thus ignoring values near the extremities of the distribution.

Interfaces: Stateful
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 center true bool Should we center the sample dataset?

Additional Methods:

Return the medians calculated by fitting the training set:

public medians() : array

Return the interquartile ranges calculated during fitting:

public iqrs() : array

Example:

use Rubix\ML\Transformers\QuartileStandardizer;

$transformer = new QuartileStandardizer(true);

Robust Standardizer

This standardizer transforms continuous features by centering them around the median and scaling by the median absolute deviation (MAD). The use of robust statistics make this standardizer more immune to outliers than the Z Scale Standardizer which used mean and variance.

Interfaces: Stateful
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 center true bool Should we center the sample dataset?

Additional Methods:

Return the medians calculated by fitting the training set:

public medians() : array

Return the median absolute deviations calculated during fitting:

public mads() : array

Example:

use Rubix\ML\Transformers\RobustStandardizer;

$transformer = new RobustStandardizer(true);

Sparse Random Projector

The Sparse Random Projector uses a random matrix sampled from a sparse Gaussian distribution (mostly 0s) to reduce the dimensionality of a dataset.

Interfaces: Stateful
Compatibility: Continuous only

Parameters:

# Param Default Type Description
1 dimensions int The number of target dimensions to project onto.

Additional Methods:

Calculate the minimum dimensionality needed given total sample size and max distortion using the Johnson-Lindenstrauss lemma:

public static minDimensions(int $n, float $maxDistortion = 0.1) : int

Example:

use Rubix\ML\Transformers\SparseRandomProjector;

$transformer = new SparseRandomProjector(30);

References:

  • D. Achlioptas. (2003). Database-friendly random projections: Johnson-Lindenstrauss with binary coins.

Stop Word Filter

Removes user-specified words from any categorical feature column including blobs of text.

Interfaces: None
Compatiblity: Categorical

Parameters:

# Param Default Type Description
1 stop words array A list of stop words to filter out of each text feature.

Additional Methods:

This transformer does not have any additional methods.

use Rubix\ML\Transformers\StopWordFilter;

$transformer = new StopWordFilter(['i', 'me', 'my', ...]);

Text Normalizer

This transformer converts all text to lowercase and optionally removes extra whitespace.

Interfaces: None
Compatibility: Categorical

Parameters:

# Param Default Type Description
1 trim false bool Should we trim excess whitespace?

Additional Methods:

This transformer does not have any additional methods.

use Rubix\ML\Transformers\TextNormalizer;

$transformer = new TextNormalizer(true);

TF-IDF Transformer

Term Frequency - Inverse Document Frequency is a measure of how important a word is to a document. The TF-IDF value increases proportionally with the number of times a word appears in a document (TF) and is offset by the frequency of the word in the corpus (IDF).

Note: This transformer assumes that its input is made up of word frequency vectors such as those created by the Word Count Vectorizer.

Interfaces: Stateful, Elastic
Compatiblity: Continuous only

Parameters:

This transformer does not have any parameters.

Additional Methods:

Return the inverse document frequencies calculated during fitting:

public idfs() : ?array

Example:

use Rubix\ML\Transformers\TfIdfTransformer;

$transformer = new TfIdfTransformer();

References:

  • S. Robertson. (2003). Understanding Inverse Document Frequency: On theoretical arguments for IDF.

Variance Threshold Filter

A type of feature selector that selects feature columns that have a greater variance than the user-specified threshold.

Interfaces: Stateful
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 threshold 0. float Feature columns with a variance greater than this threshold will be selected.

Additional Methods:

Return the columns that were selected during fitting:

public selected() : array

Example:

use Rubix\ML\Transformers\VarianceThresholdFilter;

$transformer = new VarianceThresholdFilter(50);

Word Count Vectorizer

The Word Count Vectorizer builds a vocabulary from the training samples and transforms text blobs into fixed length feature vectors. Each feature column represents a word or token from the vocabulary and the value denotes the number of times that word appears in a given sample.

Interfaces: Stateful
Compatibility: Categorical

Parameters:

# Param Default Type Description
1 max vocabulary PHP_INT_MAX int The maximum number of words to encode into each document vector.
2 min document frequency 1 int The minimum number of documents a word must appear in to be added to the vocabulary.
3 tokenizer Word object The tokenizer that extracts individual words from samples of text.

Additional Methods:

Return the fitted vocabulary i.e. the words that will be vectorized:

public vocabulary() : array

Return the size of the vocabulary:

public size() : int

Example:

use Rubix\ML\Transformers\WordCountVectorizer;
use Rubix\ML\Other\Tokenizers\SkipGram;

$transformer = new WordCountVectorizer(10000, 3, new SkipGram());

Z Scale Standardizer

A method of centering and scaling a dataset such that it has 0 mean and unit variance, also known as a Z Score.

Interfaces: Stateful, Elastic
Compatibility: Continuous

Parameters:

# Param Default Type Description
1 center true bool Should we center the sample dataset?

Additional Methods:

Return the means calculated by fitting the training set:

public means() : array

Return the variances calculated during fitting:

public variances() : array

Return the standard deviations calculated during fitting:

public stddevs() : array

Example:

use Rubix\ML\Transformers\ZScaleStandardizer;

$transformer = new ZScaleStandardizer(true);

References:

  • T. F. Chan et al. (1979). Updating Formulae and a Pairwise Algorithm for Computing Sample Variances.

Neural Network

A number of estimators in Rubix are implemented as a Neural Network under the hood. Neural nets are trained using an iterative supervised learning process called Gradient Descent with Backpropagation that repeatedly takes small steps towards minimizing a user-defined cost function. Networks can have an arbitrary number of intermediate computational layers called hidden layers. Hidden layers can perform a number of different functions including higher order feature detection, non-linear activation, normalization, and regularization.

Activation Functions

The input to a node in the network is often passed through an Activation Function (sometimes referred to as a transfer function) which determines its output behavior. In the context of a biologically inspired neural network, the activation function is an abstraction representing the rate of action potential firing of a neuron.

Activation Functions can be broken down into three classes - Sigmoidal (or S shaped) such as Hyperbolic Tangent, Rectifiers such as ELU and LeakyReLU(#leaky-relu), and Radial Basis Functions (RBFs) such as Gaussian.

ELU

Exponential Linear Units are a type of rectifier that soften the transition from non-activated to activated using the exponential function.

Parameters:

# Param Default Type Description
1 alpha 1.0 float The value at which leakage will begin to saturate. Ex. alpha = 1.0 means that the output will never be less than -1.0 when inactivated.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\ELU;

$activationFunction = new ELU(5.0);

References:

  • D. A. Clevert et al. (2016). Fast and Accurate Deep Network Learning by Exponential Linear Units.

Gaussian

The Gaussian activation function is a type of Radial Basis Function (RBF) whose activation depends only on the distance the input is from the origin.

Parameters:

This activation Function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\Gaussian;

$activationFunction = new Gaussian();

Hyperbolic Tangent

S-shaped function that squeezes the input value into an output space between -1 and 1 centered at 0.

Parameters:

This activation Function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\HyperbolicTangent;

$activationFunction = new HyperbolicTangent();

ISRU

Inverse Square Root units have a curve similar to Hyperbolic Tangent and Sigmoid but use the inverse of the square root function instead. It is purported by the authors to be computationally less complex than either of the aforementioned. In addition, ISRU allows the parameter alpha to control the range of activation such that it equals + or - 1 / sqrt(alpha).

Parameters:

# Param Default Type Description
1 alpha 1.0 float The parameter that controls the range of activation.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\ISRU;

$activationFunction = new ISRU(2.0);

References:

  • B. Carlile et al. (2017). Improving Deep Learning by Inverse Square RootvLinear Units.

Leaky ReLU

Leaky Rectified Linear Units are activation functions that output x when x > 0 or a small leakage value determined as the input times the leakage coefficient when x < 0. The amount of leakage is controlled by the leakage parameter.

Parameters:

# Param Default Type Description
1 leakage 0.1 float The amount of leakage as a ratio of the input value.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\LeakyReLU;

$activationFunction = new LeakyReLU(0.3);

References:

  • A. L. Maas et al. (2013). Rectifier Nonlinearities Improve Neural Network Acoustic Models.

ReLU

Rectified Linear Units output only the positive part of the inputs.

Note: ReLUs are analogous to half-wave rectifiers in electrical engineering.

Parameters:

This activation Function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\ReLU;

$activationFunction = new ReLU();

References:

  • V. Nair et al. (2011). Rectified Linear Units Improve RestrictedvBoltzmann Machines.

SELU

Scaled Exponential Linear Unit is a self-normalizing activation function based on the ELU activation function.

Parameters:

# Param Default Type Description
1 scale 1.05070 float The factor to scale the output by.
2 alpha 1.67326 float The value at which leakage will begin to saturate. Ex. alpha = 1.0 means that the output will never be more than -1.0 when inactivated.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\SELU;

$activationFunction = new SELU(1.05070, 1.67326);

References:

  • G. Klambauer et al. (2017). Self-Normalizing Neural Networks.

Sigmoid

A bounded S-shaped function (specifically the Logistic function) with an output value between 0 and 1.

Parameters:

This activation Function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\Sigmoid;

$activationFunction = new Sigmoid();

Softmax

The Softmax function is a generalization of the Sigmoid function that squashes each activation between 0 and 1 and all activations together add up to exactly 1.

Parameters:

# Param Default Type Description
1 epsilon 1e-8 float The smoothing parameter i.e a small value to add to the denominator for numerical stability.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\Softmax;

$activationFunction = new Softmax(1e-10);

Soft Plus

A smooth approximation of the ReLU function whose output is constrained to be positive.

Parameters:

This activation function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\SoftPlus;

$activationFunction = new SoftPlus();

References:

  • X. Glorot et al. (2011). Deep Sparse Rectifier Neural Networks.

Softsign

A function that squashes the input smoothly between -1 and 1.

Parameters:

This activation function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\Softsign;

$activationFunction = new Softsign();

References:

  • X. Glorot et al. (2010). Understanding the Difficulty of Training Deep Feedforward Neural Networks

Thresholded ReLU

Thresholded ReLU has a user-defined threshold parameter that controls the level at which the neuron is activated.

Parameters:

# Param Default Type Description
1 threshold 0. float The input value necessary to trigger an activation.

Example:

use Rubix\ML\NeuralNet\ActivationFunctions\ThresholdedReLU;

$activationFunction = new ThresholdedReLU(0.5);

References:

  • K. Konda et al. (2015). Zero-bias Autoencoders and the Benefits of Co-adapting Features.

Cost Functions

In neural networks, the cost function is a function that the network tries to minimize during training. The cost of a particular activation is defined as the difference between the output of the network and what the correct output should be given the ground truth label. Different cost functions have different ways of punishing erroneous activations and thus produce differently shaped gradients when backpropagated.

Cross Entropy

Cross Entropy, or log loss, measures the performance of a classification model whose output is a probability value between 0 and 1. Cross-entropy loss increases as the predicted probability diverges from the actual label. So predicting a probability of .012 when the actual observation label is 1 would be bad and result in a high loss value. A perfect score would have a log loss of 0.

Parameters:

This cost function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\CostFunctions\CrossEntropy;

$costFunction = new CrossEntropy();

Exponential

This cost function calculates the exponential of a prediction's squared error thus applying a large penalty to wrong predictions. The resulting gradient of the Exponential loss tends to be steeper than most other cost functions. The magnitude of the error can be scaled by the parameter tau.

Parameters:

# Param Default Type Description
1 tau 1.0 float The scaling parameter i.e. the magnitude of the error to return.

Example:

use Rubix\ML\NeuralNet\CostFunctions\Exponential;

$costFunction = new Exponential(0.5);

Huber Loss

The pseudo Huber Loss function transitions between L1 and L2 (Least Squares) loss at a given pivot point (delta) such that the function becomes more quadratic as the loss decreases. The combination of L1 and L2 loss makes Huber Loss robust to outliers while maintaining smoothness near the minimum.

Parameters:

# Param Default Type Description
1 delta 1. float The pivot point i.e the point where numbers larger will be evaluated with an L1 loss while number smaller will be evaluated with an L2 loss.

Example:

use Rubix\ML\NeuralNet\CostFunctions\HuberLoss;

$costFunction = new HuberLoss(0.5);

Least Squares

Least Squares or quadratic loss is a function that measures the squared error between the target output and the actual output of the network.

Parameters:

This cost function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\CostFunctions\LeastSquares;

$costFunction = new LeastSquares();

Relative Entropy

Relative Entropy or Kullback-Leibler divergence is a measure of how the expectation and activation of the network diverge.

Parameters:

This cost function does not have any parameters.

Example:

use Rubix\ML\NeuralNet\CostFunctions\RelativeEntropy;

$costFunction = new RelativeEntropy();

Initializers

Initializers are responsible for setting the initial weight parameters of the weight layers of a neural network. Certain activation functions respond differently when given inputs from weight layers with different initializations.

To initialize a random weight matrix:

public initialize(int $fanIn, int $fanOut) : Matrix

He

The He initializer was designed for hidden layers that feed into rectified linear unit layers such as ReLU, Leaky ReLU, and ELU. It draws from a uniform distribution with limits defined as +/- (6 / (fanIn + fanOut)) ** (1. / sqrt(2)).

Parameters:

This initializer does not have any parameters.

Example:

use Rubix\ML\NeuralNet\Initializers\He;

$initializer = new He();

References:

  • K. He et al. (2015). Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification.

Le Cun

Proposed by Yan Le Cun in a paper in 1998, this initializer was one of the first published attempts to control the variance of activations between layers through weight initialization. It remains a good default choice for many hidden layer configurations.

Parameters:

This initializer does not have any parameters.

Example:

use Rubix\ML\NeuralNet\Initializers\LeCun;

$initializer = new LeCun();

References:

  • Y. Le Cun et al. (1998). Efficient Backprop.

Normal

Generates a random weight matrix from a Gaussian distribution with user-specified standard deviation.

Parameters:

# Param Default Type Description
1 stddev 0.05 float The standard deviation of the distribution to sample from.

Example:

use Rubix\ML\NeuralNet\Initializers\Normal;

$initializer = new Normal(0.1);

Uniform

Generates a random uniform distribution centered at 0 and bounded at both ends by the parameter beta.

Parameters:

# Param Default Type Description
1 beta 0.05 float The minimum and maximum bound on the random distribution.

Example:

use Rubix\ML\NeuralNet\Initializers\Uniform;

$initializer = new Uniform(1e-3);

Xavier 1

The Xavier 1 initializer draws from a uniform distribution [-limit, limit] where limit is equal to sqrt(6 / (fanIn + fanOut)). This initializer is best suited for layers that feed into an activation layer that outputs a value between 0 and 1 such as Softmax or Sigmoid.

Parameters:

This initializer does not have any parameters.

Example:

use Rubix\ML\NeuralNet\Initializers\Xavier1;

$initializer = new Xavier1();

References:

  • X. Glorot et al. (2010). Understanding the Difficulty of Training Deep Feedforward Neural Networks.

Xavier 2

The Xavier 2 initializer draws from a uniform distribution [-limit, limit] where limit is squal to (6 / (fanIn + fanOut)) ** 0.25. This initializer is best suited for layers that feed into an activation layer that outputs values between -1 and 1 such as Hyperbolic Tangent and Softsign.

Parameters:

This initializer does not have any parameters.

Example:

use Rubix\ML\NeuralNet\Initializers\Xavier2;

$initializer = new Xavier2();

References:

  • X. Glorot et al. (2010). Understanding the Difficulty of Training Deep Feedforward Neural Networks.

Layers

Every neural network is made up of layers of computational units called neurons. Each layer processes and transforms the input from the previous layer in such a way that makes it easier for the next layer to form high-level abstractions.

There are three types of layers that form a network, Input, Hidden, and Output. A network can have as many Hidden layers as the user specifies, however, there can only be 1 Input and 1 Output layer per network.

Input Layers

The entry point for data into a neural network is the input layer which is the first layer in the network. Input layers do not have any learnable parameters.

Placeholder 1D

The Placeholder 1D input layer represents the future input values of a mini batch (matrix) of single dimensional tensors (vectors) to the neural network.

Parameters:

# Param Default Type Description
1 inputs None int The number of inputs to the neural network.

Example:

use Rubix\ML\NeuralNet\Layers\Placeholder1D;

$layer = new Placeholder1D(100);

Hidden Layers

In multilayer networks, hidden layers are responsible for transforming the input space in such a way that can be linearly separable by the final output layer.

Activation

Activation layers apply a nonlinear activation function to their inputs.

Parameters:

# Param Default Type Description
1 activation fn None object The function computes the activation of the layer.

Example:

use Rubix\ML\NeuralNet\Layers\Activation;
use Rubix\ML\NeuralNet\ActivationFunctions\ReLU;

$layer = new Activation(new ReLU());

Alpha Dropout

Alpha Dropout is a type of dropout layer that maintains the mean and variance of the original inputs in order to ensure the self-normalizing property of SELU networks with dropout. Alpha Dropout fits with SELU networks by randomly setting activations to the negative saturation value of the activation function at a given ratio each pass.

Note: Alpha Dropout is generally only used in the context of SELU networks. Use regular Dropout for other types of neural nets.

Parameters:

# Param Default Type Description
1 ratio 0.1 float The ratio of neurons that are dropped during each training pass.

Example:

use Rubix\ML\NeuralNet\Layers\AlphaDropout;

$layer = new AlphaDropout(0.1);

References:

  • G. Klambauer et al. (2017). Self-Normalizing Neural Networks.

Batch Norm

Normalize the activations of the previous layer such that the mean activation is close to 0 and the activation standard deviation is close to 1. Batch Norm can be used to reduce the amount of covariate shift within the network making it possible to use higher learning rates and converge faster under some circumstances.

Parameters:

This layer does not have any parameters.

Example:

use Rubix\ML\NeuralNet\Layers\BatchNorm;

$layer = new BatchNorm();

References:

  • S. Ioffe et al. (2015). Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift.
  • T. F. Chan et al. (1979). Updating Formulae and a Pairwise Algorithm for Computing Sample Variances.

Dense

Dense layers are fully connected neuronal layers, meaning each neuron is connected to each other in the previous layer by a weighted synapse. The majority of the parameters in a standard feedforward network are usually contained within the Dense hidden layers of the network.

Parameters:

# Param Default Type Description
1 neurons None int The number of neurons in the layer.
2 initializer He object The random weight initializer to use.

Example:

use Rubix\ML\NeuralNet\Layers\Dense;
use Rubix\ML\NeuralNet\Initializers\He;

$layer = new Dense(100, new He());

Dropout

Dropout layers temporarily disable neurons during each training pass. Dropout is a regularization and model averaging technique for reducing overfitting in neural networks by preventing complex co-adaptations on training data.

Parameters:

# Param Default Type Description
1 ratio 0.5 float The ratio of neurons that are dropped during each training pass.

Example:

use Rubix\ML\NeuralNet\Layers\Dropout;

$layer = new Dropout(0.5);

References:

  • N. Srivastava et al. (2014). Dropout: A Simple Way to Prevent Neural Networks from Overfitting.

Noise

This layer adds random Gaussian noise to the inputs to the layer with a standard deviation given as a parameter. Noise added to neural network activations acts as a regularizer by indirectly adding a penalty to the weights through the cost function in the output layer.

Parameters:

# Param Default Type Description
1 stddev 0.1 float The standard deviation of the gaussian noise to add to the inputs.

Example:

use Rubix\ML\NeuralNet\Layers\Noise;

$layer = new Noise(2.0);

References:

  • C. Gulcehre et al. (2016). Noisy Activation Functions.

PReLU

The PReLU layer uses leaky ReLU activation functions whose leakage coefficients are parameterized and optimized on a per neuron basis during training.

Parameters:

# Param Default Type Description
1 initial 0.25 float The value to initialize the alpha (leakage) parameters with.

Example:

use Rubix\ML\NeuralNet\Layers\PReLU;

$layer = new PReLU(0.1);

References:

  • K. He et al. (2015). Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification.

Output Layers

Activations are read directly from the Output layer when making predictions. The type of output layer will determine the type of Estimator the network can bestow (i.e Binary Classifier, Multiclass Classifier, or Regressor).

Binary

The Binary layer consists of a single Sigmoid neuron capable of distinguishing between two discrete classes. The Binary layer is useful for neural networks that output a binary class prediction such as yes or no.

Parameters:

# Param Default Type Description
1 classes None array The unique class labels of the binary classification problem.
2 alpha 1e-4 float The L2 regularization penalty.
3 cost fn Cross Entropy object The function that penalizes the activities of bad predictions.

Example:

use Rubix\ML\NeuralNet\Layers\Binary;
use Rubix\ML\NeuralNet\CostFunctions\CrossEntropy;

$layer = new Binary(['yes', 'no'], 1e-3, new CrossEntropy());

Continuous

The Continuous output layer consists of a single linear neuron that outputs a scalar value useful for regression problems.

Parameters:

# Param Default Type Description
1 alpha 1e-4 float The L2 regularization penalty.
2 cost fn Least Squares object The function that penalizes the activities of bad predictions.

Example:

use Rubix\ML\NeuralNet\Layers\Continuous;
use Rubix\ML\NeuralNet\CostFunctions\HuberLoss;

$layer = new Continuous(1e-5, new HuberLoss(3.0));

Multiclass

The Multiclass output layer gives a joint probability estimate of a multiclass classification problem using the Softmax activation function.

Parameters:

# Param Default Type Description
1 classes None array The unique class labels of the multiclass classification problem.
2 alpha 1e-4 float The L2 regularization penalty.
3 cost fn Cross Entropy object The function that penalizes the activities of bad predictions.

Example:

use Rubix\ML\NeuralNet\Layers\Multiclass;
use Rubix\ML\NeuralNet\CostFunctions\RelativeEntropy;

$layer = new Multiclass(['yes', 'no', 'maybe'], 1e-4, new RelativeEntropy());

Optimizers

Gradient Descent is an algorithm that takes iterative steps towards finding the best set of weights in a neural network. Rubix provides a number of pluggable Gradient Descent optimizers that control the step of each parameter in the network.

AdaGrad

Short for Adaptive Gradient, the AdaGrad Optimizer speeds up the learning of parameters that do not change often and slows down the learning of parameters that do enjoy heavy activity.

Parameters:

# Param Default Type Description
1 rate 0.01 float The learning rate. i.e. the master step size.

Example:

use Rubix\ML\NeuralNet\Optimizers\AdaGrad;

$optimizer = new AdaGrad(0.125);

References:

  • J. Duchi et al. (2011). Adaptive Subgradient Methods for Online Learning and Stochastic Optimization.

Adam

Short for Adaptive Moment Estimation, the Adam Optimizer combines both Momentum and RMS prop to achieve a balance of velocity and stability. In addition to storing an exponentially decaying average of past squared gradients like RMSprop, Adam also keeps an exponentially decaying average of past gradients, similar to Momentum. Whereas Momentum can be seen as a ball running down a slope, Adam behaves like a heavy ball with friction, which thus prefers flat minima in the error surface.

Parameters:

# Param Default Type Description
1 rate 0.001 float The learning rate. i.e. the master step size.
2 momentum 0.9 float The decay rate of the Momentum property.
3 rms 0.999 float The decay rate of the RMS property.

Example:

use Rubix\ML\NeuralNet\Optimizers\Adam;

$optimizer = new Adam(0.0001, 0.9, 0.999);

References:

  • D. P. Kingma et al. (2014). Adam: A Method for Stochastic Optimization.

Cyclical

The Cyclical optimizer uses a global learning rate that cycles between the lower and upper bound over a designated period while also decaying the upper bound by a factor of gamma each step. Cyclical learning rates have been shown to help escape local minima and saddle points thus achieving higher accuracy.

Parameters:

# Param Default Type Description
1 lower 0.001 float The lower bound on the learning rate.
2 upper 0.006 float The upper bound on the learning rate.
3 steps 100 int The number of steps in every half cycle.
4 decay 0.99994 float The exponential decay factor to decrease the learning rate by every step.

Example:

use Rubix\ML\NeuralNet\Optimizers\Cyclical;

$optimizer = new Cyclical(0.001, 0.005, 1000);

References:

  • L. N. Smith. (2017). Cyclical Learning Rates for Training Neural Networks.

Momentum

Momentum adds velocity to each step until exhausted. It does so by accumulating momentum from past updates and adding a factor of the previous velocity to the current step.

Parameters:

# Param Default Type Description
1 rate 0.001 float The learning rate. i.e. the master step size.
2 decay 0.9 float The Momentum decay rate.

Example:

use Rubix\ML\NeuralNet\Optimizers\Momentum;

$optimizer = new Momentum(0.001, 0.925);

References:

  • D. E. Rumelhart et al. (1988). Learning representations by back-propagating errors.

RMS Prop

An adaptive gradient technique that divides the current gradient over a rolling window of magnitudes of recent gradients.

Parameters:

# Param Default Type Description
1 rate 0.001 float The learning rate. i.e. the master step size.
2 decay 0.9 float The RMS decay rate.

Example:

use Rubix\ML\NeuralNet\Optimizers\RMSProp;

$optimizer = new RMSProp(0.01, 0.9);

References:

  • T. Tieleman et al. (2012). Lecture 6e rmsprop: Divide the gradient by a running average of its recent magnitude.

Step Decay

A learning rate decay optimizer that reduces the learning rate by a factor of the decay parameter whenever it reaches a new floor. The number of steps needed to reach a new floor is defined by the steps parameter.

Parameters:

# Param Default Type Description
1 rate 0.01 float The learning rate. i.e. the master step size.
2 steps 100 int The size of every floor in steps. i.e. the number of steps to take before applying another factor of decay.
3 decay 1e-3 float The decay factor to decrease the learning rate by every floor.

Example:

use Rubix\ML\NeuralNet\Optimizers\StepDecay;

$optimizer = new StepDecay(0.1, 50, 1e-3);

Stochastic

A constant learning rate optimizer based on the original Stochastic Gradient Descent paper.

Parameters:

# Param Default Type Description
1 rate 0.001 float The learning rate. i.e. the step size.

Example:

use Rubix\ML\NeuralNet\Optimizers\Stochastic;

$optimizer = new Stochastic(0.001);

Kernels

Distance

Distance kernels measure the distance between points in vector space. They are used throughout Rubix in Estimators that employ the concept of distance to make predictions such as K Nearest Neighbors, K Means, and Local Outlier Factor.

Note: Distance is only defined for continuous data points in Rubix.

Canberra

A weighted version of Manhattan distance which computes the L1 distance between two coordinates in a vector space.

Parameters:

This kernel does not have any parameters.

Example:

use Rubix\ML\Kernels\Distance\Canberra;

$kernel = new Canberra();

Cosine

Cosine Similarity is a measure that ignores the magnitude of the distance between two vectors thus acting as strictly a judgement of orientation. Two vectors with the same orientation have a cosine similarity of 1, two vectors oriented at 90° relative to each other have a similarity of 0, and two vectors diametrically opposed have a similarity of -1. To be used as a distance function, we subtract the Cosine Similarity from 1 in order to satisfy the positive semi-definite condition, therefore the Cosine distance is a number between 0 and 2.

Parameters:

This kernel does not have any parameters.

Example:

use Rubix\ML\Kernels\Distance\Cosine;

$kernel = new Cosine();

Diagonal

The Diagonal (sometimes called Chebyshev) distance is a measure that constrains movement to horizontal, vertical, and diagonal from a point. An example that uses Diagonal movement is a chess board.

Parameters:

This kernel does not have any parameters.

Example:

use Rubix\ML\Kernels\Distance\Diagonal;

$kernel = new Diagonal();

Euclidean

This is the ordinary straight line (bee line) distance between two points in Euclidean space. The associated norm of the Euclidean distance is called the L2 norm.

Parameters:

This kernel does not have any parameters.

Example:

use Rubix\ML\Kernels\Distance\Euclidean;

$kernel = new Euclidean();

Jaccard

This generalized Jaccard distance is a measure of similarity that one sample has to another with a range from 0 to 1. The higher the percentage, the more dissimilar they are.

Parameters:

This kernel does not have any parameters.

Example:

use Rubix\ML\Kernels\Distance\Jaccard;

$kernel = new Jaccard();

Manhattan

A distance metric that constrains movement to horizontal and vertical, similar to navigating the city blocks of Manhattan. An example that used this type of movement is a checkers board.

Parameters:

This kernel does not have any parameters.

Example:

use Rubix\ML\Kernels\Distance\Manhattan;

$kernel = new Manhattan();

Minkowski

The Minkowski distance is a metric in a normed vector space which can be considered as a generalization of both the Euclidean and Manhattan distances. When the lambda parameter is set to 1 or 2, the distance is equivalent to Manhattan and Euclidean respectively.

Parameters:

# Param Default Type Description
1 lambda 3.0 float Controls the curvature of the unit circle drawn from a point at a fixed distance.

Example:

use Rubix\ML\Kernels\Distance\Minkowski;

$kernel = new Minkowski(4.0);

SVM

Support Vector Machine kernels are used in the context of SVM-based estimators to project sample vectors into a non-linear feature space, allowing them to marginalize non-linear data.

Linear

A simple linear kernel computed by the dot product of two vectors.

Parameters:

This kernel does not have any parameters.

Example:

use Rubix\ML\Kernels\SVM\Linear;

$kernel = new Linear();

Polynomial

This kernel projects a sample vector using polynomials of the p'th degree.

Parameters:

# Param Default Type Description
1 degree 3 int The degree of the polynomial.
2 gamma null float The kernel coefficient.
3 coef0 0. float The independent term.

Example:

use Rubix\ML\Kernels\SVM\Polynomial;

$kernel = new Polynomial(3, null, 0.);

RBF

Non linear radial basis function computes the distance from a centroid or origin.

Parameters:

# Param Default Type Description
1 gamma null float The kernel coefficient.

Example:

use Rubix\ML\Kernels\SVM\RBF;

$kernel = new RBF(null);

Sigmoidal

S shaped nonliearity kernel with output values ranging from -1 to 1.

Parameters:

# Param Default Type Description
1 gamma null float The kernel coefficient.
2 coef0 0. float The independent term.

Example:

use Rubix\ML\Kernels\SVM\Sigmoidal;

$kernel = new Sigmoidal(null, 0.);

Cross Validation

Cross validation is the process of testing the generalization performance of a model.

Validators

Validators take an Estimator instance, Labeled Dataset object, and validation Metric and return a validation score that measures the generalization performance of the model using one of various cross validation techniques. There is no need to train the Estimator beforehand as the Validator will automatically train it on subsets of the dataset created by the testing algorithm.

public test(Estimator $estimator, Labeled $dataset, Validation $metric) : float

Return the validation scores computed at last test time:

public scores() : ?array

Example:

use Rubix\ML\CrossValidation\KFold;
use Rubix\ML\CrossValidation\Metrics\Accuracy;

$validator = new KFold(10);

$score = $validator->test($estimator, $dataset, new Accuracy());

var_dump($score);

Output:

float(0.869)

Hold Out

Hold Out is a simple cross validation technique that uses a hold out validation set. The advantages of Hold Out is that it is quick, but it doesn't allow the model to train on the entire training set.

Parameters:

# Param Default Type Description
1 ratio 0.2 float The ratio of samples to hold out for testing.
2 stratify false bool Should we stratify the dataset before splitting?

Example:

use Rubix\ML\CrossValidation\HoldOut;

$validator = new HoldOut(0.25, true);

K Fold

K Fold is a technique that splits the training set into K individual sets and for each training round uses 1 of the folds to measure the validation performance of the model. The score is then averaged over K. For example, a K value of 10 will train and test 10 versions of the model using a different testing set each time.

Parameters:

# Param Default Type Description
1 k 10 int The number of times to split the training set into equal sized folds.
2 stratify false bool Should we stratify the dataset before folding?

Example:

use Rubix\ML\CrossValidation\KFold;

$validator = new KFold(5, true);

Leave P Out

Leave P Out tests the model with a unique holdout set of P samples for each round until all samples have been tested.

Note: Leave P Out can become slow with large datasets and small values of P.

Parameters:

# Param Default Type Description
1 p 10 int The number of samples to leave out each round for testing.

Example:

use Rubix\ML\CrossValidation\LeavePOut;

$validator = new LeavePOut(50);

Monte Carlo

Repeated Random Subsampling or Monte Carlo cross validation is a technique that takes the average validation score over a user-supplied number of simulations (randomized splits of the dataset).

Parameters:

# Param Default Type Description
1 simulations 10 int The number of simulations to run i.e the number of tests to average.
2 ratio 0.2 float The ratio of samples to hold out for testing.
3 stratify false bool Should we stratify the dataset before splitting?

Example:

use Rubix\ML\CrossValidation\MonteCarlo;

$validator = new MonteCarlo(30, 0.1);

Validation Metrics

Validation metrics are for evaluating the performance of an Estimator given the ground truth labels.

Compute a validation score, pass in the predictions from an estimator along with the ground-truth labels:

public score(array $predictions, array $labels) : float

Output the range of values the validation score can take on in a 2-tuple:

public range() : array

Return a list of estimators that metric is compatible with:

public compatibility() : array

Example:

use Rubix\ML\CrossValidation\Metrics\MeanAbsoluteError;

$metric = new MeanAbsoluteError();

$score = $metric->score($predictions, $labels);

var_dump($metric->range());

var_dump($score);

Output:

array(2) {
  [0]=> float(-INF)
  [1]=> int(0)
}

float(-0.99846070553066)

Accuracy

Accuracy is a quick classification and anomaly detection metric defined as the number of true positives over all samples in the testing set.

Compatibility: Classification, Anomaly Detection
Range: 0 to 1

Example:

use Rubix\ML\CrossValidation\Metrics\Accuracy;

$metric = new Accuracy();

Completeness

A ground truth clustering metric that measures the ratio of samples in a class that are also members of the same cluster. A cluster is said to be complete when all the samples in a class are contained in a cluster.

Compatibility: Clustering
Range: 0 to 1

Example:

use Rubix\ML\CrossValidation\Metrics\Completeness;

$metric = new Completeness();

F1 Score

A weighted average of precision and recall with equal relative contribution.

Compatibility: Classification, Anomaly Detection
Range: 0 to 1

Example:

use Rubix\ML\CrossValidation\Metrics\F1Score;

$metric = new F1Score();

Homogeneity

A ground truth clustering metric that measures the ratio of samples in a cluster that are also members of the same class. A cluster is said to be homogeneous when the entire cluster is comprised of a single class of samples.

Compatibility: Clustering
Range: 0 to 1

Example:

use Rubix\ML\CrossValidation\Metrics\Homogeneity;

$metric = new Homogeneity();

Informedness

Informedness is a measure of the probability that an estimator will make an informed decision. The index was suggested by W.J. Youden as a way of summarizing the performance of a diagnostic test. Its value ranges from 0 through 1 and has a zero value when the test gives the same proportion of positive results for groups with and without the disease, i.e the test is useless.

Compatibility: Classification, Anomaly Detection
Range: 0 to 1

Example:

use Rubix\ML\CrossValidation\Metrics\Informedness;

$metric = new Informedness();

MCC

Matthews Correlation Coefficient measures the quality of a classification. It takes into account true and false positives and negatives and is generally regarded as a balanced measure which can be used even if the classes are of very different sizes. The MCC is in essence a correlation coefficient between the observed and predicted binary classifications; it returns a value between −1 and +1. A coefficient of +1 represents a perfect prediction, 0 no better than random prediction and −1 indicates total disagreement between prediction and observation.p

Compatibility: Classification, Anomaly Detection
Range: -1 to 1

Example:

use Rubix\ML\CrossValidation\Metrics\MCC;

$metric = new MCC();

Mean Absolute Error

A metric that measures the average amount that a prediction is off by given some ground truth (labels).

Compatibility: Regression
Range: -∞ to 0

Example:

use Rubix\ML\CrossValidation\Metrics\MeanAbsoluteError;

$metric = new MeanAbsoluteError();

Mean Squared Error

A regression metric that punishes bad predictions the worse they get by averaging the squared error over the testing set.

Compatibility: Regression
Range: -∞ to 0

Example:

use Rubix\ML\CrossValidation\Metrics\MeanSquaredError;

$metric = new MeanSquaredError();

Median Absolute Error

Median Absolute Error (MAE) is a robust measure of the error that ignores highly erroneous predictions.

Compatibility: Regression
Range: -∞ to 0

Example:

use Rubix\ML\CrossValidation\Metrics\MedianAbsoluteError;

$metric = new MedianAbsoluteError();

Rand Index

The Adjusted Rand Index is a measure of similarity between the clustering and some ground truth that is adjusted for chance. It considers all pairs of samples that are assigned in the same or different clusters in the predicted and empirical clusterings.

Compatibility: Regression
Range: -1 to 1

Example:

use Rubix\ML\CrossValidation\Metrics\RandIndex;

$metric = new RandIndex();

References:

  • W. M. Rand. (1971). Objective Criteria for the Evaluation of Clustering Methods.

RMS Error

Root Mean Squared (RMS) Error or average L2 loss is a metric that is used to compute the residuals of a regression problem.

Compatibility: Regression
Range: -∞ to 0

Example:

use Rubix\ML\CrossValidation\Metrics\RMSError;

$metric = new RMSError();

R Squared

The coefficient of determination or R Squared (R²) is the proportion of the variance in the dependent variable that is predictable from the independent variable(s).

Compatibility: Regression
Range: -∞ to 1

Example:

use Rubix\ML\CrossValidation\Metrics\RSquared;

$metric = new RSquared();

V Measure

V Measure is the harmonic balance between homogeneity and completeness and is used as a measure to determine the quality of a clustering.

Compatibility: Clustering
Range: 0 to 1

Example:

use Rubix\ML\CrossValidation\Metrics\VMeasure;

$metric = new VMeasure();

References:

  • A. Rosenberg et al. (2007). V-Measure: A conditional entropy-based external cluster evaluation measure.

Reports

Reports offer a comprehensive view of the performance of an estimator given the problem in question.

To generate a report from the predictions of an estimator given some ground truth labels:

public generate(array $predictions, array $labels) : array

Return a list of estimators that report is compatible with:

public compatibility() : array

Example:

use Rubix\ML\Reports\ConfusionMatrix;

$report = new ConfusionMatrix(['positive', 'negative']);

$result = $report->generate($predictions, $labels);

Aggregate Report

A report that aggregates the results of multiple reports. The reports are indexed by the key given at construction time.

Parameters:

# Param Default Type Description
1 reports array An array of report objects to aggregate.

Example:

use Rubix\ML\CrossValidation\Reports\AggregateReport;
use Rubix\ML\CrossValidation\Reports\ConfusionMatrix;
use Rubix\ML\CrossValidation\Reports\MulticlassBreakdown;

$report = new AggregateReport([
	'breakdown' => new MulticlassBreakdown(),
	'matrix1' => new ConfusionMatrix(['wolf', 'lamb']),
	'matrix2' => new ConfusionMatrix(['human', 'gorilla']),
]);

$result = $report->generate($estimator, $testing);

Confusion Matrix

A Confusion Matrix is a table that visualizes the true positives, false, positives, true negatives, and false negatives of a classifier. The name stems from the fact that the matrix makes it easy to see the classes that the classifier might be confusing.

Compatibility: Classification, Anomaly Detection

Parameters:

# Param Default Type Description
1 classes All array The classes to compare in the matrix.

Example:

use Rubix\ML\CrossValidation\Reports\ConfusionMatrix;

$report = new ConfusionMatrix(['dog', 'cat', 'turtle']);

$result = $report->generate($estimator, $testing);

var_dump($result);

Output:

  array(3) {
    ["dog"]=> array(2) {
      ["dog"]=> int(842)
      ["cat"]=> int(5)
      ["turtle"]=> int(0)
    }
    ["cat"]=>
    array(2) {
      ["dog"]=> int(0)
      ["cat"]=> int(783)
      ["turtle"]=> int(3)
    }
    ["turtle"]=>
    array(2) {
      ["dog"]=> int(31)
      ["cat"]=> int(79)
      ["turtle"]=> int(496)
    }
  }

Contingency Table

A Contingency Table is used to display the frequency distribution of class labels among a clustering of samples.

Compatibility: Clustering

Parameters:

This report does not have any parameters.

Example:

use Rubix\ML\CrossValidation\Reports\ContingencyTable;

$report = new ContingencyTable();

$result = $report->generate($estimator, $testing);

var_dump($result);

Output:

array(3) {
    [1]=>
    array(3) {
      [1]=> int(13)
      [2]=> int(0)
      [3]=> int(2)
    }
    [2]=>
    array(3) {
      [1]=> int(1)
      [2]=> int(0)
      [3]=> int(12)
    }
    [0]=>
    array(3) {
      [1]=> int(0)
      [2]=> int(14)
      [3]=> int(0)
    }
  }

Multiclass Breakdown

A report that drills down in to each unique class outcome. The report includes metrics such as Accuracy, F1 Score, MCC, Precision, Recall, Fall Out, and Miss Rate.

Compatibility: Classification, Anomaly Detection

Parameters:

# Param Default Type Description
1 classes All array The classes to break down.

Example:

use Rubix\ML\CrossValidation\Reports\MulticlassBreakdown;

$report = new MulticlassBreakdown(['wolf', 'lamb']);

$result = $report->generate($estimator, $testing);

var_dump($result);

Output:

["label"]=> array(2) {
	["wolf"]=> array(19) {
      	["accuracy"]=> float(0.6)
      	["precision"]=> float(0.66666666666667)
      	["recall"]=> float(0.66666666666667)
      	["specificity"]=> float(0.5)
      	["negative_predictive_value"]=> float(0.5)
      	["false_discovery_rate"]=> float(0.33333333333333)
      	["miss_rate"]=> float(0.33333333333333)
      	["fall_out"]=> float(0.5)
      	["false_omission_rate"]=> float(0.5)
     	["f1_score"]=> float(0.66666666666667)
      	["mcc"]=> float(0.16666666666667)
      	["informedness"]=> float(0.16666666666667)
      	["markedness"]=> float(0.16666666666667)
      	["true_positives"]=> int(2)
      	["true_negatives"]=> int(1)
      	["false_positives"]=> int(1)
      	["false_negatives"]=> int(1)
      	["cardinality"]=> int(3)
      	["density"]=> float(0.6)
    }

Residual Analysis

Residual Analysis is a Report that measures the differences between the predicted and actual values of a regression problem in detail.

Compatibility: Regression

Parameters:

This report does not have any parameters.

Example:

use Rubix\ML\CrossValidation\Reports\ResidualAnaysis;

$report = new ResidualAnalysis();

$result = $report->generate($estimator, $testing);

var_dump($result);

Output:

  array(12) {
    ["mean_absolute_error"]=> float(0.44927554249285)
    ["median_absolute_error"]=> float(0.30273889978541)
    ["mean_squared_error"]=> float(0.44278193357447)
    ["rms_error"]=> float(0.66541861529001)
	["mean_squared_log_error"]=> float(-0.35381010755)
	["r_squared"]=> float(0.99393263320234)
    ["error_mean"]=> float(0.14748941084881)
    ["error_variance"]=> float(0.42102880726195)
    ["error_skewness"]=> float(-2.7901397847317)
    ["error_kurtosis"]=> float(12.967400285518)
    ["error_min"]=> float(-3.5540079974946)
    ["error_max"]=> float(1.4097829828182)
    ["cardinality"]=> int(80)
  }

Generators

Dataset generators produce synthetic data of a user-specified shape, dimensionality, and cardinality. Synthetic data is useful for augmenting a dataset or for quick testing and demonstration purposes.

To generate a Dataset object with n samples (rows):

public generate(int $n) : Dataset

Return the dimensionality of the samples produced by the generator:

public dimensions() : int

Example:

use Rubix\ML\Datasets\Generators\Blob;

$generator = new Blob([0, 0], 1.0);

$dataset = $generator->generate(3);

var_dump($generator->dimensions());

var_dump($dataset->samples());

Output:

int(2)

object(Rubix\ML\Datasets\Unlabeled)#24136 (1) {
  ["samples":protected]=>
  array(3) {
    [0]=>
    array(2) {
      [0]=> float(-0.2729673885539)
      [1]=> float(0.43761840244204)
    }
    [1]=>
    array(2) {
      [0]=> float(-1.2718092282012)
      [1]=> float(-1.9558245484829)
    }
    [2]=>
    array(2) {
      [0]=> float(1.1774185431405)
      [1]=> float(0.05168623824664)
    }
  }
}

Agglomerate

An Agglomerate is a collection of other generators each given a label. Agglomerates are useful for classification, clustering, and anomaly detection problems where the label is a discrete value.

Data: Continuous
Label: Categorical

Parameters:

# Param Default Type Description
1 generators array A collection of generators keyed by their user-specified label (0 indexed by default).
2 weights Auto array A set of arbitrary weight values corresponding to a generator's contribution to the agglomeration.

Additional Methods:

Return the normalized weights of each generator in the agglomerate:

public weights() : array

Example:

use Rubix\ML\Datasets\Generators\Agglomerate;

$generator = new Agglomerate([
	new Blob([5, 2], 1.0),
	new HalfMoon([-3, 5], 1.5, 90.0, 0.1),
	new Circle([2, -4], 2.0, 0.05),
], [
	5, 6, 3, // An arbitrary set of weights
]);

Blob

A normally distributed n-dimensional blob of samples centered at a given mean vector. The standard deviation can be set for the whole blob or for each feature column independently. When a global value is used, the resulting blob will be isotropic.

Data: Continuous
Label: None

Parameters:

# Param Default Type Description
1 center [0.0, 0.0] array The coordinates of the center of the blob i.e. a centroid vector.
2 stddev 1.0 float or array Either the global standard deviation or an array with the SD for each feature column.

Additional Methods:

This generator does not have any additional methods.

Example:

use Rubix\ML\Datasets\Generators\Blob;

$generator = new Blob([-1.2, -5.0, 2.6, 0.8], 0.25);

Circle

Creates a circle of points in 2 dimensions.

Data: Continuous
Label: Continuous

Parameters:

# Param Default Type Description
1 x 0.0 float The x coordinate of the center of the circle.
2 y 0.0 float The y coordinate of the center of the circle.
3 scale 1.0 float The scaling factor of the circle.
4 noise 0.1 float The amount of Gaussian noise to add to each data point as a ratio of the scaling factor.

Additional Methods:

This generator does not have any additional methods.

Example:

use Rubix\ML\Datasets\Generators\Circle;

$generator = new Circle(0.0, 0.0, 100, 0.1);

Half Moon

Generate a dataset consisting of 2 dimensional samples that form a half moon shape when plotted.

Data: Continuous
Label: Continuous

Parameters:

# Param Default Type Description
1 x 0.0 float The x coordinate of the center of the half moon.
2 y 0.0 float The y coordinate of the center of the half moon.
3 scale 1.0 float The scaling factor of the half moon.
4 rotate 90.0 float The amount in degrees to rotate the half moon counterclockwise.
5 noise 0.1 float The amount of Gaussian noise to add to each data point as a percentage of the scaling factor.

Additional Methods:

This generator does not have any additional methods.

Example:

use Rubix\ML\Datasets\Generators\HalfMoon;

$generator = new HalfMoon(4.0, 0.0, 6, 180.0, 0.2);

Swiss Roll

Generate a 3-dimensional swiss roll dataset with continuous valued labels. The labels are the inputs to the swiss roll transformation and are suitable for non-linear regression problems.

Data: Continuous
Label: Continuous

Parameters:

# Param Default Type Description
1 x 0.0 float The x coordinate of the center of the swiss roll.
2 y 0.0 float The y coordinate of the center of the swiss roll.
3 z 0.0 float The z coordinate of the center of the swiss roll.
4 scale 1.0 float The scaling factor of the swiss roll.
5 depth 21.0 float The depth of the swiss roll i.e the scale of the y axis.
6 noise 0.3 float The standard deviation of the gaussian noise.

Additional Methods:

This generator does not have any additional methods.

Example:

use Rubix\ML\Datasets\Generators\SwissRoll;

$generator = new SwissRoll(5.5, 1.5, -2.0, 10, 21.0, 0.2);

Other

This section includes all classes that do not fall under a specific category.

Guessing Strategies

Guesses can be thought of as a type of weak prediction. Unlike a real prediction, guesses are made using limited information. A guessing Strategy attempts to use such information to formulate an educated guess. Guessing is utilized in both Dummy Estimators (Dummy Classifier, Dummy Regressor) as well as the Missing Data Imputer.

The Strategy interface provides an API similar to Transformers as far as fitting, however, instead of being fit to an entire dataset, each Strategy is fit to an array of either continuous or discrete values.

To fit a Strategy to an array of values:

public fit(array $values) : void

To make a guess based on the fitted data:

public guess() : mixed

Blurry Percentile

A strategy that guesses within the domain of the p-th percentile of the fitted data plus some gaussian noise.

Continuous

Parameters:

# Param Default Type Description
1 p 50.0 float The index of the percentile to predict where 50 is the median.
2 blur 0.1 float The amount of gaussian noise to add to the guess as a factor of the median absolute deviation (MAD).

Example:

use Rubix\ML\Other\Strategies\BlurryPercentile;

$strategy = new BlurryPercentile(34.0, 0.2);

Constant

Always guess a constant value.

Continuous

Parameters:

# Param Default Type Description
1 value 0. float The value to guess.

Example:

use Rubix\ML\Other\Strategies\Constant;

$strategy = new Constant(17.);

K Most Frequent

This strategy outputs one of K most frequent discrete values at random.

Categorical

Parameters:

# Param Default Type Description
1 k 1 int The number of most frequency categories to consider when formulating a guess.

Example:

use Rubix\ML\Other\Strategies\KMostFrequent;

$strategy = new KMostFrequent(5);

Lottery

Hold a lottery in which each category has an equal chance of being picked.

Categorical

Parameters:

This strategy does not have any parameters.

Example:

use Rubix\ML\Other\Strategies\Lottery;

$strategy = new Lottery();

Mean

This strategy always predicts the mean of the fitted data.

Continuous

Parameters:

This strategy does not have any parameters.

Example:

use Rubix\ML\Other\Strategies\Mean;

$strategy = new Mean();

Popularity Contest

Hold a popularity contest where the probability of winning (being guessed) is based on the category's prior probability.

Categorical

Parameters:

This strategy does not have any parameters.

Example:

use Rubix\ML\Other\Strategies\Lottery;

$strategy = new PopularityContest();

Wild Guess

It is what you think it is. Make a guess somewhere in between the minimum and maximum values observed during fitting with equal probability given to all values within range.

Continuous

Parameters:

# Param Default Type Description
1 precision 2 int The number of decimal places of precision for each guess.

Example:

use Rubix\ML\Other\Strategies\WildGuess;

$strategy = new WildGuess(5);

Helpers

Data Type

Determine the data type of a variable according to Rubix ML's type system.

To determine the data type of a variable:

public determine($variable) : int

Note: The return value is an integer encoding of the datatype defined as constants on the DataType class.

Return true if the variable is categorical:

public isCategorical($variable) : bool

Return true if the variable is categorical:

public isContinuous($variable) : bool

Return true if the variable is a PHP resource:

public isResource($variable) : bool

Return true if the variable is an unrecognized data type:

public isOther($variable) : bool

Example:

use Rubix\ML\Other\Helpers\DataType;

var_dump(DataType::determine('string'));

var_dump(DataType::isContinuous(16));

var_dump(DataType::isCategorical(18));

Output:

int(2) // Categorical
bool(true)
bool(false)

Params

Generate distributions of values to use in conjunction with Grid Search or other forms of model selection and/or cross validation.

To generate a unique distribution of integer parameters:

public sta