The source code and dataset for ICDM2020 paper: Dynamic Graph Collaborative Filtering, implemented in Pytorch.
Our code mainly mainly refers to jodie: https://github.com/srijankr/jodie/, data download method, running environment and initialization are the same as jodie.
If you make use of this code or the DGCF algorithm in your work, please cite the following paper:
@inproceedings{li2020dynamic,
title={Dynamic graph collaborative filtering},
author={Li, Xiaohan and Zhang, Mengqi and Wu, Shu and Liu, Zheng and Wang, Liang and Philip, S Yu},
booktitle={2020 IEEE International Conference on Data Mining (ICDM)},
pages={322--331},
year={2020},
organization={IEEE}
}
To train the DGCF model using the data/<network>.csv dataset, use the following command. This will save a model for every epoch in the saved_models/<network>/ directory.
python DGCF.py --network <network> --model DGCF --epochs 50 --method attention --adj
This code can be given the following command-line arguments:
--network: choose to the train data:reddit\wikipedia\lastfm
--model: this is the name of the model
--epochs: this is the maximum number of interactions to train the model.
--embedding_dim: this is the number of dimensions of the dynamic embedding.
--method: this is the type of aggregator function in second-order aggregation
--adj: this is a boolean input indicating if use the second update.
--length： this is the aggregator size in second-order aggegator function.
To evaluate the performance of the interaction prediction task in one epoch, use the following command:
python evaluate_interaction_prediction.py --network --model --method
To evaluate the performance of the interaction prediction task in all epoch, user the followinig command:
python evaluate_all.py --network --model --method
For detailed code execution, you can refer to the command line in the ./Shell folder.