Awesome

SMIN

SMIN

Source code for Social Recommendation with Self-Supervised Metagraph Informax Network

Requirements

• Pytorch(1.5.0)
• DGL(0.4.3), installation-https://github.com/dmlc/dgl

More Details

data preprocessing

• CiaoDVD

  rating.mat and trust.mat as original data source from https://www.cse.msu.edu/~tangjili/datasetcode/truststudy.htm

  loadMat.py: training/test data partition

  run ./dataset/CiaoDVD/loadMat.py to perform preprocessing

  GenerateMetaPath.py: metapath generation

  run ./dataset/CiaoDVD/GenrateMetaPath.py to perform generation process

  GenerateSubGraph.py: generate k-hop subfigures for Informax module

  run ./dataset/CiaoDVD/GenerateSubGraph.py to perform k-hop subfigure construction

• Similar data preprocessing steps are applied in Epinions and Yelp data.

Code running example

Run main.py：

python main.py --dataset CiaoDVD --hide_dim 16 --layer_dim [16] --lr 0.05 --reg 0.05  --lambda1 0.06 --lambda2 0.002 

Combination of sub-modules and code organization

Interface

BPRData.py: for generating the positive and negative instances corresponding to training and test set, respectively

evaluate.py: perform evaluation of our proposed framework

MV_MIL (Multi-view Graph-Structured Mutual Information Learning Paradigm)

informax.py: incorporate the learned social- and knowledge-aware dependence to guide the user-item interaction embedding process through deriving mutual information terms from different views.

gcn.py and graphconv.py: the basic graph neural network architecture with the convolutional relation encoder

ToolScripts

TimeLogger.py: log timestamp information

tools.py: convert the sparse matrices to sparse tensors

model.py

model class integrates the graph neural network architecture with high-order relation modeling SemanticAttention class defines the attention mechanism to aggregate metapath-specific representations

main.py In the trainModel of Hope class, we adopt the model.py to optimize the loss of user-item interaction learing component.

The joint learning component of i) meta-relation heterogeneity encoding and ii) multi-view graph-structured mutual information learning is defined informax.py.