【摘要】 现有的多行为推荐模型忽略了不同行为之间存在的优化不平衡问题。为解决这一问题，提出了一种融入自注意力和对比学习的多行为推荐模型(multi-behavior recommendation integrating self-attention and contrastive learning, SACL)。首先，根据用户与商品的交互行为类型构建独立的交互视图，通过图神经网络挖掘用户与物品之间的关联关系，提取用户的不同行为特征和兴趣偏好特征。其次，在行为间与用户间进行对比学习，捕捉不同行为下的相同用户特征，增强辅助行为信息的利用率。然后，基于自注意力机制设计出一个多行为优化模块，根据用户的多行为特征以及对比学习特征设计定义不同的编码方式，生成具有行为依赖关系的元知识；设计了一个自注意多行为损失权重网络，根据元知识平衡不同行为的训练损失权重，从而区分对目标行为的影响差异并降低辅助行为噪声。提出的模型在Tmall和IJCAI-Contest数据集上进行的实验表明，相较于最优基线DPT,SACL的命中率(HR)平均提升了10%,归一化折损率(NDCG)平均提升了14%,验证了SACL模型对平衡优化多行为推荐任务的有效性。更多还原

【Abstract】 Existing multi-behavior recommendation models ignore the optimization imbalance problem that exists between different behaviors. To solve this problem, this paper proposed a multi-behavior recommendation model integrating self-attention and contrastive learning(SACL). Firstly, it constructed independent interaction views based on different types of user-item interaction behaviors, and explored correlation relationships between users and items through graph neural networks to extract different behavior characteristics and interest preference features of users. Secondly, it performed contrastive learning between behaviors and users to capture the same user characteristics under different behaviors and enhanced the utilization of auxiliary behavior information. Then, it designed a multi-behavior optimization module based on the self-attention mechanism which defined different encoding methods based on the multi-behavior features and contrastive learning features of users to generate meta-knowledge with behavior dependencies. It designed a self-attentive multi-behavior loss weighting network to balance the training loss weights of different behaviors based on meta-knowledge, thus distinguishing the differences in the impacts on target behavior and reducing the auxiliary behavior noises. Experiments on the Tmall dataset and the IJCAI-Contest dataset show that the proposed model improves the hit rate(HR) of SACL by an average of 10% and the normalized discount rate(NDCG) by an average of 14% compared to the optimal baseline DPT, which verifies the effectiveness of SACL model for balanced optimization of multi-behavior recommendation tasks.更多还原