Personalized Multi-Behavior Sequential Recommendation with Temporal-Aware Contrastive Learning and Self-Attention

Sequential recommendation (SR) is designed to capture sequential patterns and a user’s dynamic preferences from their historical interactions. Traditional sequential recommendation only considers a single behavior type of users, but in actual recommendation scenarios, users tend to produce multiple types of behaviors (e.g., page view, Add-to-favorite, purchase), so multi-behavior sequential recommendation (MBSR) is proposed. This paper aims to address the challenges posed by MBSR. First, users exhibit diverse and dynamic multi-behavior patterns due to different personal habits. Second, there are complex synergies between different behavior types and between behaviors and projects. In addition, existing MBSR methods do not fully consider the influence of time context. To address these challenges, we propose a Personalized Multi-Behavior Sequential Recommendation Framework (PMB) with Temporal-Aware Contrastive Learning and Self-Attention. Specifically, PMB first developed a personalized behavior awareness generator at the presentation layer to extract dynamic discriminant behavior patterns, then improved the self-attention mechanism through Sequential behavior aware collaborative extractor, and finally used time perception contrast learning. Used to model and learn vertical temporal proximity to project interactions between users. Finally, we conducted experiments on four benchmark datasets, demonstrating the effectiveness of PMB by significantly improving recommendation performance compared to various baseline models.