/ComiRec

Source code and dataset for KDD 2020 paper "Controllable Multi-Interest Framework for Recommendation"

Primary LanguagePython

Controllable Multi-Interest Framework for Recommendation

Original implementation for paper Controllable Multi-Interest Framework for Recommendation.

Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, Jie Tang

Accepted to KDD 2020 ADS Track!

Prerequisites

  • Python 3
  • TensorFlow-GPU >= 1.8 (< 2.0)
  • Faiss-GPU

Getting Started

Installation

Dataset

Training

Training on the existing datasets

You can use python src/train.py --dataset {dataset_name} --model_type {model_name} to train a specific model on a dataset. Other hyperparameters can be found in the code. (If you share the server with others or you want to use the specific GPU(s), you may need to set CUDA_VISIBLE_DEVICES.)

For example, you can use python src/train.py --dataset book --model_type ComiRec-SA to train ComiRec-SA model on Book dataset.

When training a ComiRec-DR model, you should set --learning_rate 0.005.

Training on your own datasets

If you want to train models on your own dataset, you should prepare the following three(or four) files:

  • train/valid/test file: Each line represents an interaction, which contains three numbers <user_id>,<item_id>,<time_stamp>.
  • category file (optional): Each line contains two numbers <item_id>,<cate_id> used for computing diversity..

Common Issues

The computation of NDCG score.
I'm so sorry that the computation of NDCG score in the original version (now in `paper` branch) is not consistent with the definition in the paper, as mentioned in the issue #6. I have updated the computation of NDCG score in the `master` branch according to the correct definition. For reproducing the NDCG scores reported in the paper, please use the `paper` branch. By the way, I personally recommend to use the reported results of recall and hit rate only.

If you have ANY difficulties to get things working in the above steps, feel free to open an issue. You can expect a reply within 24 hours.

Acknowledgement

The structure of our code is based on MIMN.

Cite

Please cite our paper if you find this code useful for your research:

@inproceedings{cen2020controllable,
  title = {Controllable Multi-Interest Framework for Recommendation},
  author = {Cen, Yukuo and Zhang, Jianwei and Zou, Xu and Zhou, Chang and Yang, Hongxia and Tang, Jie},
  booktitle = {Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
  year = {2020},
  pages = {2942–2951},
  publisher = {ACM},
}