Simplified self-supervised learning for hybrid propagation graph-based recommendation.

Recent progress in Graph Convolutional Networks (GCNs) has facilitated their extensive application in recommendation, yielding notable performance gains. Nevertheless, existing GCN-based recommendation approaches are confronted with several challenges: (1) how to effectively leverage multi-order graph connectivity to derive meaningful node embeddings; (2) faced with sparse raw data, how to augment supervision signals without relying on auxiliary information; (3) given that GCNs necessitate the aggregation of neighborhood nodes, and the sparsity of these nodes can exacerbate the impact of noise data, how to mitigate the noise problem inherent in the raw data. For tackling aforementioned challenges, we devise a new hybrid propagation GCN-based method named S3HGN, incorporating a simplified self-supervised learning paradigm for recommendation. Specifically, we introduce the concept of nonlinear propagation into the common linear GCN framework to explore both low-order and high-order hybrid connectivity relationships, and perform residual prediction through weighted summation. Furthermore, we design a simplified self-supervised learning strategy to construct auxiliary tasks, which creates contrastive views by directly applying dropout operation twice to the final representation, enhancing the model's robustness against noise data. Extensive experiments on three publicly available datasets, using eight representative graph-based collaborative filtering models, confirm the effectiveness and robustness of the proposed S3HGN.