Inhomogeneous Social Recommendation with Hypergraph Convolutional Networks

This is our Pytorch implementation for the paper:

Zirui Zhu, Chen Gao, Xu Chen, Nian Li, Depeng Jin, and Yong Li. Inhomogeneous Social Recommendation with Hypergraph Convolutional Networks, in ICDE 2022.

Introduction

Social HyperGraph Convolutionl Network (SHGCN) is a new recommendation framework based on hypergraph convolution, effectively utilizing the triple social relations.

Environment Requirement

The code has been tested under Python 3.6.10. The required packages are as follows:

• Pytorch == 1.6.0
• numpy == 1.19.1
• scipy == 1.5.2
• pandas == 1.1.1

Example to Run the Codes

• Beidian dataset

python main.py --dataset Beidian --model SHGCN --gpu 0 --emb_dim 32 --num_layer 2 --epoch 500 --batch_size 4096

• Beibei dataset

python main.py --dataset Beibei --model SHGCN --gpu 0 --emb_dim 32 --num_layer 2 --epoch 500 --batch_size 4096

Dataset

There are two datasets released here. Each contains four files.

• data.train

  • Training set.
  • Each line contains a purchase log, which can be represented as:
    • (user ID, item ID)

• data.val

  • Validation set.
  • Each line contains a purchase log, which can be represented as:
    • (user ID, item ID)

• data.test

  • Testing set.
  • Each line contains a purchase log, which can be represented as:
    • (user ID, item ID)

• social.share

  • Social interactions logs.
  • Each line contains a triplet. It can be represented uniformly as
    • (user1 ID, user2 ID, item ID)
  • In the Beidian dataset, each triplet represents a social sharing behavior.
  • In the Beibei dataset, each triplet represents a group buying behavior.