LightGCN

LightGCN started from a heretical observation: in collaborative filtering, the standard GCN machinery may be doing more harm than good. User and item IDs have no meaningful input features, so feature transformation matrices and nonlinear activations mostly add parameters to overfit. LightGCN removes them entirely. The model keeps only neighborhood aggregation on the user-item bipartite graph, propagates ID embeddings through a few linear layers, and combines the embeddings from all layers with a weighted sum. Training is the classic BPR pairwise ranking loss. The result is simpler, faster, and easier to train than NGCF, the GCN-based recommender it set out to beat.\n\nThe SIGIR 2020 paper reports recall@20 of 0.1830 on Gowalla, 0.0649 on Yelp2018 and 0.0411 on Amazon-Book, beating NGCF by 10-25% relative on every dataset, and the numbers have been independently reproduced (OpenHGNN's replication matches the paper to the third decimal). LightGCN became one of the most influential recommender-system models of its era and the backbone that later contrastive methods like SGL built on.

Paper-reported recall@20: 0.1830 Gowalla, 0.0649 Yelp2018, 0.0411 Amazon-Book - 10-25% relative gains over NGCF (He et al., SIGIR 2020). Independently reproduced by OpenHGNN (0.1841/0.0648/0.0414).