snudatalab / popcon

PopCon

This project is a PyTorch implementation of Aggregately Diversified Bundle Recommendation via Popularity Debiasing and Configuration-aware Reranking (PopCon), which is published in PAKDD 2023.

Overview

The overview of PopCon is as follows. PopCon consists of two phases, model training phase and reranking phase. In the training phase, PopCon trains a bundle recommendation model such as DAM or CrossCBR as a backbone while mitigating its popularity bias by a popularity-based negative sampling. In the raranking phase, PopCon selects candidate bundles for each user and reranks the candidates by a configuration-aware reranking algorithm to maximize both accuracy and aggregate diversity. For more details, please refer to our paper.

Prerequisties

Our implementation is based on Python 3.8 and Pytorch 1.8.1. Please see the full list of packages required to our codes in requirements.txt.

Datasets

We use 3 datasets in our work: Steam, Youshu, and NetEase. We include the preprocessed datasets in the repository: data/{data_name}.

Backbone model

We provide DAM, one of the state-of-the-art bundle recommendation models, as a backbone. It is defined in models.py. CrossCBR, another state-of-the-art model, is available at https://github.com/mysbupt/CrossCBR

Running the code

You can run the pretraining code by python pretrain.py with arguments --epochs and --alpha. You can also run the reranking code by python reranking.py with arguments --beta and --n. To run reranking.py, running pretrain.py must precede because it returns a recommendation results of a model. We provide demo.sh, which reproduces the experiments of our work.

Citation

Please cite this paper when you use our code.

@inproceedings{conf/pakdd/JeonKLLK23,
  author    = {Hyunsik Jeon and
               Jongjin Kim and
               Jaeri Lee and
               Jong-eun Lee and
               U Kang},
  title     = {Aggregately Diversified Bundle Recommendation via Popularity Debiasing and Configuration-aware Reranking},
  booktitle = {PAKDD},
  year      = {2023},
}

License

This software may be used only for non-commercial purposes (e.g., research evaluation) in universities. Please contact Prof. U Kang ([email protected]) if you want to use it for other purposes or use it in places other than universities.

popcon's Issues

about the neg.pkl file

I could not find the method that you used to generate neg.pkl file in either the paper or the slide. Can you please publish the method or the code that you used to get the file? Thank you so much.