Self-supervised Graph Learning for Recommendation

Representation learning on user-item graph for recommendation has evolved\nfrom using single ID or interaction history to exploiting higher-order\nneighbors. This leads to the success of graph convolution networks (GCNs) for\nrecommendation such as PinSage and LightGCN. Despite effectiveness, we argue\nthat they suffer from two limitations: (1) high-degree nodes exert larger\nimpact on the representation learning, deteriorating the recommendations of\nlow-degree (long-tail) items; and (2) representations are vulnerable to noisy\ninteractions, as the neighborhood aggregation scheme further enlarges the\nimpact of observed edges.\n In this work, we explore self-supervised learning on user-item graph, so as\nto improve the accuracy and robustness of GCNs for recommendation. The idea is\nto supplement the classical supervised task of recommendation with an auxiliary\nself-supervised task, which reinforces node representation learning via\nself-discrimination. Specifically, we generate multiple views of a node,\nmaximizing the agreement between different views of the same node compared to\nthat of other nodes. We devise three operators to generate the views -- node\ndropout, edge dropout, and random walk -- that change the graph structure in\ndifferent manners. We term this new learning paradigm as\n\\textit{Self-supervised Graph Learning} (SGL), implementing it on the\nstate-of-the-art model LightGCN. Through theoretical analyses, we find that SGL\nhas the ability of automatically mining hard negatives. Empirical studies on\nthree benchmark datasets demonstrate the effectiveness of SGL, which improves\nthe recommendation accuracy, especially on long-tail items, and the robustness\nagainst interaction noises. Our implementations are available at\n\\url{https://github.com/wujcan/SGL}.\n