RecSys Benchmark

This is the source code provided for the paper A Thorough Performance Benchmarking on Lightweight Embedding-based Recommender Systems (Link).

If you find this repo helpful, please give a star and cite the below paper if possible:

@article{tran2024thorough,
  title={A Thorough Performance Benchmarking on Lightweight Embedding-based Recommender Systems},
  author={Tran, Hung Vinh and Chen, Tong and Nguyen, Quoc Viet Hung and Huang, Zi and Cui, Lizhen and Yin, Hongzhi},
  journal={arXiv preprint arXiv:2406.17335},
  year={2024}
}

Installation

# clone this repo
git clone [email protected]:chenxing1999/recsys-benchmark.git

# cd to folder
cd recsys-benchmark

# Create virtual env
python -m venv env
source env/bin/activate

# install
pip install -e '.[dev]'

Quickstart

Train SGL-WA:

python scripts/lightgcn/train_lightgcn.py configs/yelp2018/base_config.yaml

Run Hyperparam search with Optuna

mkdir -p checkpoints/lightgcn/

# Storing logs at "logs/sgl-wa" and best checkpoint at "checkpoints/lightgcn/sgl-wa.pth".
python scripts/lightgcn/exp_find_hparams.py -c configs/yelp2018/base_config.yaml -l logs/sgl-wa -p checkpoints/lightgcn/sgl-wa.pth

Train DeepFM:

python scripts/deepfm/train_deepfm.py configs/criteo/base_config.yaml

Run Hyperparam search with Optuna

mkdir -p checkpoints/deepfm/

# Storing logs at "logs/deepfm" and best checkpoint at "checkpoints/deepfm/original.pth".
python scripts/deepfm/exp_find_hparams.py -c configs/criteo/base_config.yaml -l logs/deepfm -p checkpoints/deepfm/original.pth

For more detail with LightGCN, click here

For more detail with DeepFM, click here. DCN-Mix uses the same API with DeepFM.

Other artifacts

Please send an email to [email protected] for necessary checkpoints and data splits.

Acknowledge

This source code is based on:

https://github.com/gusye1234/LightGCN-PyTorch
https://github.com/rixwew/pytorch-fm