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 extended journal version is published in TORS.

• Paper: 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:

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.