README.md

Inductive Matrix Completion Using Graph Autoencoder

Paper link: https://arxiv.org/pdf/2108.11124.pdf

Author's code: https://github.com/swtheing/IMC-GAE

The implementation does not handle side-channel features and mini-epoching and thus achieves slightly worse performance when using node features.

Credit: Wei shen (@swtheing)

Requirements

Latest tested combination: Python 3.8.1 + PyTorch 1.4.0 + DGL 0.5.2.

Install PyTorch

Install DGL

Other required python libraries: numpy, scipy, pandas, h5py, networkx, tqdm ,bidict etc.

Data

Supported datasets: ml-100k, ml-1m, ml-10m, flixster, douban, yahoo_music

How to run

To train on Flixster, type:

    python -u train.py --data_name=flixster \
                       --use_one_hot_fea \
                       --gcn_agg_accum=sum \
                       --device 0 \
                       --ARR 0.00000000000 \
                       --train_early_stopping_patience 200 \
                       --layers 2 \
                       --gcn_agg_units 30 \
                       --train_lr 0.01 \
                       --data_valid_ratio 0.1 \
                       --model_activation tanh \
                       --gcn_out_units 30

Results: RMSE=0.8838

To train on douban, type:

    python -u train.py --data_name=douban \
                       --use_one_hot_fea \
                       --gcn_agg_accum=sum \
                       --device 0 \
                       --ARR 0.00000000000 \
                       --train_early_stopping_patience 1800 \
                       --layers 5 \
                       --gcn_agg_units 90 \
                       --train_lr 0.01 \
                       --data_valid_ratio 0.05 \
                       --model_activation tanh \
                       --gcn_out_units 80

Results: RMSE=0.7211

To train on yahoo_music, type:

    python -u train.py --data_name=yahoo_music \
                       --use_one_hot_fea \
                       --gcn_agg_accum=sum \
                       --device 0 \
                       --ARR 0.00000000000 \
                       --train_early_stopping_patience 20 \
                       --layers 3 \
                       --gcn_agg_units 30 \
                       --train_lr 0.01 \
                       --data_valid_ratio 0.1 \
                       --model_activation tanh \
                       --gcn_out_units 30

Results: RMSE=18.7018

To train on MovieLens-100K, type:

    python -u train.py --data_name=ml-100k \
                       --use_one_hot_fea \
                       --device 0 \
                       --ARR 0.00004 \
                       --layers 2 \
                       --data_valid_ratio 0.05 \
                       --model_activation tanh 

Results: RMSE=0.8967

To train on MovieLens-1M, type:

    python -u train.py --data_name=ml-1m \
                       --use_one_hot_fea \
                       --device 0 \
                       --ARR 0.000004 \
                       --layers 2 \
                       --data_valid_ratio 0.05 \
                       --model_activation tanh 

Results: RMSE=0.8290

Reference

If you find the code useful, please cite our paper.

@article{shen2021inductive,
  title={Inductive Matrix Completion Using Graph Autoencoder},
  author={Shen, Wei and Zhang, Chuheng and Tian, Yun and Zeng, Liang and He, Xiaonan and Dou, Wanchun and Xu, Xiaolong},
  journal={arXiv preprint arXiv:2108.11124},
  year={2021}
}

9/17/2021