【摘要】 现有自监督社交推荐模型大多通过人工启发式图增强和单一关系视图间对比的策略构建自监督信号，性能受到增强自监督信号质量的影响，难以自适应地抑制噪声.由此，文中提出基于自监督图掩码神经网络的社交推荐模型.首先，分别构建用户社交和物品分类的单一关系视图及高阶连通异构图，采用图掩码学习范式指导用户社交图进行自适应和可学习的数据增强.然后，设计异构图编码器，学习视图中的潜在语义，跨视图对用户、物品嵌入进行对比学习，完成自监督任务，分别对用户、物品嵌入进行加权融合，完成推荐任务.最后，利用多任务训练策略联合优化自监督学习任务、推荐任务和图掩码任务.在3个真实数据集上的实验表明文中模型性能具有一定提升.更多还原

【Abstract】 The existing self-supervised social recommendation models mostly construct self-supervised signals through the strategies of manual heuristic graph enhancement and contrasts between single-relational views. Thus, the performance of the model is easily affected by the quality of enhanced self-supervised signals, making it challenging to adaptively suppress noise. To solve these problems, a social recommendation model based on self-supervised graph masked neural networks(SGMN) is proposed. Firstly, single-relational views for user-social interaction and item classification are constructed respectively, as well as high-order connected heterogeneous graphs. Specifically, the graph masked learning paradigm is adopted to guide adaptive and learnable data augmentation for the user-social graph. Secondly, a heterogeneous graph encoder is designed to learn the latent semantics of the views, and cross-view contrastive learning is performed on user and item embeddings to complete self-supervised tasks. Then, weighted fusion is conducted on the user and item embeddings separately for recommendation task. Finally, a multi-task training strategy is employed to jointly optimize self-supervised learning, recommendation and graph masked tasks. Experiments on three real datasets demonstrate a certain performance improvement of SGMN.更多还原