Contents

• Acknowledgement
• Citing This Paper
• Environment Setup
• Running Framework
• Datasets

Acknowledgement

This repository contains the source code for the Graph Transfer Learning project developed by the Northeastern University's SPIRAL research group. This research was generously supported by the National Science Foundation (grants IIS-1741197, CCF-1750539) and Google via GCP credit support.

Citing This Paper

Please cite the following paper if you intend to use this code for your research.

A. Gritsenko, Y. Guo, K. Shayestehfard, A. Moharrer, J. Dy, S. Ioannidis, "Graph Transfer Learning", ICDM, 2021.

Environment Setup

Please install the python dependencies found in requirements.txt with:

pip install -r requirements.txt

Running Framework

To fully address the generic nature of the algorithm introduced in the original paper, we provide a fully-customizable framework with a wide variety of parameters for node embedding, model creation and training.

The following arguments can be specified to train node embeddings:

  --nembedding          Size of the output embedding vector
  --topology_similarity Similarity measure between nodes of the same graph in
                        graph topological space
  --embedding_type      Type of embedding function: skipgram, unified
  -embedding_similarity Similarity measures between nodes of the same graph in
                        embedding space
  --nwalks              Number of node2vec random walks
  --walk_length         Length of random walk
  --window_size         Width of sliding window in random walks
  --p                   Parameter p for node2vec random walks
  --q                   Parameter q for node2vec random walks
  --nnegative           Number of negative samples used in skip-gram
  --scale_negative      Specifies whether to scale outputs for negative
                        samples
  --graph_distance      Pairwise distance measure between nodes in the
                        embedding space (matrix D)

The following arguments can be specified to create and train model:

  --similarity_loss     Loss function between similarities in topological and
                        embedding spaces for nodes of the same graph
  --depth               Number of hidden layers in Prediction Branch
  --activation_function Activation function for Prediction Branch neurons
  --prediction_loss     Loss function for Prediction Branch
  --transfer_mode       Specifies transfer learning mode
  --alpha               Weight of graph matching loss
  --beta                Specifies whether to scale parts of P-optimization
                        loss
  --learning_rate       Learning rate
  --batch_size          Number of instances in each batch
  --epochs              Number of epochs
  --early_stopping      Number of epochs with no improvement after which
                        training will be stopped. If <=0, no early stopping is
                        used

For a full list of arguments to run a framework, you may use --help.

Datasets

All datasets referenced in the original paper are presented in the folder data. A user can run the framework on either provided datasets, or any arbitrary ones by specifying the dataset folder via --load_path and --dataset parameters. The framework expects the following files to be present in the specified dataset directory:

• a GraphA.txt file containing graph's A adjacency matrix,
• a GraphALabels_cluster.txt file containing class labels for each graph A node,
• a GraphALabels_infection.txt file containing infection labels for each graph A node. Optionally, a dataset directory can contain GraphATrain.txt and GraphATest.txt files containing node indices for train and test splits, respectively. If these files are not provided, graph nodes are split randomly into train and test subsets with a ratio 8:2. We provide GraphATrain.txt and GraphATest.txt files for all real-world datasets for the reproducibility purposes. Additionally, we provide original dataset files for each real-world graph.

The following real-world datasets are presented in the folder data:

• Zachary Karate Club

  W. W. Zachary, “An information flow model for conflict andfission in small groups”, Journal of Anthropological Research, 1977

• Email

  J. Leskovecet et al., “Graph evolution: Densification andshrinking diameters”, ACM TKDD, 2007

• Infectious Disease Transmission Dataset

  M. Salathé et al., “A high-resolution human contact networkfor infectious disease transmission”, PNAS, 2010

For the details on synthetic dataset construction, please refer to Section V.A of the original paper.