MGDCF

Source code (TensorFlow) and dataset of the paper "MGDCF: Distance Learning via Markov Graph Diffusion for Neural Collaborative Filtering", which is accepted by IEEE Transactions on Knowledge and Data Engineering (TKDE).

This repository contains the TensorFlow implementation of our paper. The official PyTorch implementation of 'MGDCF: Distance Learning via Markov Graph Diffusion for Neural Collaborative Filtering' is available in a separate repository, which can be accessed here: https://github.com/CrawlScript/Torch-MGDCF.

Implementations and Paper Links

• PyTorch Implementation: Torch-MGDCF
• TensorFlow Implementation: TensorFlow-MGDCF
• Paper Access:
  • IEEE Xplore: https://ieeexplore.ieee.org/document/10384729
  • ArXiv: https://arxiv.org/abs/2204.02338

InfoBPR Loss

We propose a simple yet powerful InfoBPR Loss for ranking. We also build an out-of-the-box library for InfoBPR:

• https://github.com/CrawlScript/InfoBPR

The InfoBPR support both TensorFlow and PyTorch, and it can be installed with pip.

Requirements

• Linux
• Python 3.7
• tensorflow == 2.7.0
• tf_geometric == 0.1.5
• tf_sparse == 0.0.17
• grecx >= 0.0.6
• tqdm=4.51.0

Run MGDCF

You can run MGDCF with the following command:

cd scripts/gnn_speed/${DATASET}
sh $SCRIPT_NAME

For example, if you want to run Hetero-MGDCF on yelp, the command is:

cd scripts/gnn_speed/yelp
sh run_gdcf_HeteroMGDCF_yelp.sh

Note that the parameter settings are in the shell scripts, and you should only modify the "gpu_ids" argument.

Cite

@ARTICLE{10384729,
  author={Jun Hu and Bryan Hooi and Shengsheng Qian and Quan Fang and Changsheng Xu},
  journal={IEEE Transactions on Knowledge and Data Engineering}, 
  title={MGDCF: Distance Learning via Markov Graph Diffusion for Neural Collaborative Filtering}, 
  year={2024},
  volume={},
  number={},
  pages={1-16},
  doi={10.1109/TKDE.2023.3348537}
}