Source code and datasets for "Revisiting LightGCN: Unexpected Inflexibility, Inconsistency, and A Remedy Towards Improved Recommendation" (RecSys 2024 Short Paper).
Paper: link
Supplementary document: link
The source code of LightGCN++ can be found here.
- To run LightGCN++ with the specific configuration for each dataset, simply run:
./run.sh
- For the version supporting Intel Gaudi devices:
./run_gaudi.sh
- To run with different
$\alpha$ ,$\beta$ , and$\gamma$ , run:
python main.py --dataset=[DATASET NAME] --alpha [ALPHA VALUE] --beta [BETA VALUE] --gamma [GAMMA VALUE]
e.g.,
python main.py --dataset="yelp2018" --alpha 0.6 --beta -0.1 --gamma 0.1
- By default, we recommend using
$\alpha=0.6$ ,$\beta=-0.1$ , and$\gamma=0.2$ .
We used five datasets: LastFM, MovieLens, Gowalla, Yelp, and Amazon. You can find them here.
We also provide a code that runs in the SELFRec framework. You can find it here.
- To run LightGCN++ with the specific configuration for each dataset, simply run:
./run.sh [DATASET]
e.g., to run lastfm with its optimal hyperparameters:
./run.sh lastfm
The optimal hyperparameter configurations of each dataset are as follows:
| Dataset | LastFM | CiteULike | MovieLens-1M | Gowalla | Yelp | Amazon-Sports | Amazon-Beauty | Amazon-Book | MovieLens-10M | Alibaba |
|---|---|---|---|---|---|---|---|---|---|---|
| α | 0.6 | 0.5 | 0.4 | 0.6 | 0.6 | 0.6 | 0.5 | 0.6 | 0.6 | 0.6 |
| β | -0.1 | -0.1 | 0.1 | -0.1 | -0.1 | -0.1 | 0.0 | -0.1 | -0.1 | 0.0 |
| γ | 0.0 | 0.4 | 0.0 | 0.2 | 0.0 | 0.0 | 0.2 | 0.2 | 0.0 | 0.1 |
This code is implemented based on the open source LightGCN PyTorch code. This research was supported in part by the NAVER-Intel Co-Lab. The work was conducted by KAIST and reviewed by both NAVER and Intel.