RecBole-GNN v0.1.0

RecBole-GNN v0.1.0 Release Notes

Bingo! After a long period of effort, we finally develop RecBole-GNN, a recommendation library tailored for graph neural network (GNN) based recommendation algorithms. RecBole-GNN is an extended library of RecBole, which is built upon PyTorch and PyG and designed for reproducing and developing recommendation algorithms based on GNNs.

In this initial release, we partially refactored RecBole for graph-structured data and implemented several popular GNN-based recommendation algorithms. In addition, we also published a leaderboard for reference. More details will be introduced in the following part:

• Highlights
• Implemented Model
• Leaderboard

RecBole-GNN is still in its rapid development period, we warmly welcome any type of PRs, including new models, bug reports, and suggestions.

Highlights

• Easy-to-use and unified API: Our library shares unified API and input (atomic files) as RecBole.
• Efficient and reusable graph processing: We provide highly efficient and reusable basic datasets, dataloaders and layers for graph processing and learning.
• Extensive graph library: Graph neural networks from widely-used libraries like PyG are incorporated. Recently proposed graph algorithms can be easily equipped and compared with existing methods.

Implemented Model

Our library currently includes three major categories of models:

• General Recommendation with user-item interaction graphs: NGCF (#6), LightGCN (#2), SGL (#4), HMLET (#15), NCL (#5);
• Sequential Recommendation with session/sequence graphs: SR-GNN (#3), GC-SAN (#3), NISER+ (#3), LESSR (#7), TAGNN (#13), GCE-GNN (#17), SGNN-HN (#25);
• Social Recommendation with social networks: DiffNet (#8), MHCN (#27), SEPT (#19).

Leaderboard

We carefully tune the hyper-parameters of the implemented models on the mainstream datasets of each research field and release leaderboards, respectively:

• The leaderboard of GNN-based general recommendation algorithms on MovieLens-1M dataset;
• The leaderboard of GNN-based sequential recommendation algorithms on Diginetica dataset;
• The leaderboard of GNN-based social recommendation algorithms on LastFM dataset;

Acknowledgement

Many thanks to the great efforts contributed by Yupeng @hyp1231, Lanling @Sherry-XLL, and Changxin @ChangxinTian. The team members come from RUC AI Box, which is supported by Prof. Wayne Xin Zhao. It's hoped that Recbole-GNN serves as an important step towards RecBole Community.