A PyTorch implementation of "Signed Graph Convolutional Network" (ICDM 2018).
Due to the fact much of today's data can be represented as graphs, there has been a demand for generalizing neural network models for graph data. One recent direction that has shown fruitful results, and therefore growing interest, is the usage of graph convolutional neural networks (GCNs). They have been shown to provide a significant improvement on a wide range of tasks in network analysis, one of which being node representation learning. The task of learning low-dimensional node representations has shown to increase performance on a plethora of other tasks from link prediction and node classification, to community detection and visualization. Simultaneously, signed networks (or graphs having both positive and negative links) have become ubiquitous with the growing popularity of social media. However, since previous GCN models have primarily focused on unsigned networks (or graphs consisting of only positive links), it is unclear how they could be applied to signed networks due to the challenges presented by negative links. The primary challenges are based on negative links having not only a different semantic meaning as compared to positive links, but their principles are inherently different and they form complex relations with positive links. Therefore we propose a dedicated and principled effort that utilizes balance theory to correctly aggregate and propagate the information across layers of a signed GCN model. We perform empirical experiments comparing our proposed signed GCN against state-of-the-art baselines for learning node representations in signed networks. More specifically, our experiments are performed on four real-world datasets for the classical link sign prediction problem that is commonly used as the benchmark for signed network embeddings algorithms.
This repository provides an implementation for SGCN as described in the paper:
Signed Graph Convolutional Network. Tyler Derr, Yao Ma, and Jiliang Tang ICDM, 2018. [Paper]
The codebase is implemented in Python 3.5.2. package versions used for development are just below.
networkx 1.11 tqdm 4.28.1 numpy 1.15.4 pandas 0.23.4 texttable 1.5.0 scipy 1.1.0 argparse 1.1.0 sklearn 0.20.0 torch 0.4.1.post2 torch-geometric 0.3.1 torchvision 0.2.1
The code takes an input graph in a csv file. Every row indicates an edge between two nodes separated by a comma. The first row is a header. Nodes should be indexed starting with 0. Sample graphs for the
Bitcoin Alpha and
Bitcoin OTC graphs are included in the
input/ directory. The structure of the edge dataset is the following:
|NODE ID 1||NODE ID 2||Sign|
An attributed dataset for an
Erdos-Renyi graph is also included in the input folder. The node feature dataset rows are sorted by ID increasing. The structure of the features csv has to be the following:
|Feature 1||Feature 2||Feature 3||...||Feature d|
Learning an embedding is handled by the
src/main.py script which provides the following command line arguments.
Input and output options
--edge-path STR Input graph path. Default is `input/bitcoin_otc.csv`. --features-path STR Features path. Default is `input/bitcoin_otc.csv`. --embedding-path STR Embedding path. Default is `output/embedding/bitcoin_otc_sgcn.csv`. --regression-weights-path STR Regression weights path. Default is `output/weights/bitcoin_otc_sgcn.csv`. --log-path STR Log path. Default is `logs/bitcoin_otc_logs.json`.
--epochs INT Number of SGCN training epochs. Default is 100. --reduction-iterations INT Number of SVD epochs. Default is 128. --reduction-dimensions INT SVD dimensions. Default is 30. --seed INT Random seed value. Default is 42. --lamb FLOAT Embedding regularization parameter. Default is 1.0. --test-size FLOAT Test ratio. Default is 0.2. --learning-rate FLOAT Learning rate. Default is 0.001. --weight-decay FLOAT Weight decay. Default is 10^-5. --layers LST Layer sizes in model. Default is [64, 32]. --spectral-features BOOL Layer sizes in autoencoder model. Default is True --general-features BOOL Loss calculation for the model. Sets spectral features to False.
The following commands learn a node embedding, regression weights and write the embedding to disk. The node representations are ordered by the ID. The layer sizes can be set manually.
Training an SGCN model on the default dataset. Saving the embedding, regression weights and logs at default paths.
Creating an SGCN model of the default dataset with a 96-64-32 architecture.
python src/main.py --layers 96 64 32
Creating a single layer SGCN model with 32 features.
python src/main.py --layers 32
Creating a model with some custom learning rate and epoch number.
python src/main.py --learning-rate 0.001 --epochs 200
Training a model on another dataset with features present - a signed
Erdos-Renyi graph. Saving the weights, output and logs in a custom folder.
python src/main.py --general-features --edge-path input/erdos_renyi_edges.csv --features-path input/erdos_renyi_features.csv --embedding-path output/embedding/erdos_renyi.csv --regression-weights-path output/weights/erdos_renyi.csv --log-path logs/erdos_renyi.json