Decision Transformer based Target-aware Feature Selection Network for Behavior-oriented Recommendation

Learning dynamic user preferences has become an important component for E-commerce platforms to make sequential recommendations. Many works have been focused on modeling users’ purchase behavior for sequential recommendations, while less work has considered users’ dynamic behavioral intentions. With this concern, we introduce a new recommendation problem, Behavior-Oriented Recommendation (BOR), where recommenders must retrieve items based on users’ diverse behavioral intentions. In this paper, We figure out that due to the discrepancy between BOR and traditional scenarios, existing recommendation models are struggling in discovering preference drift due to changes in behavioral intentions. To tackle this problem, we propose a Target-aware Feature Selection Network (TFSN). More specifically, a sequence encoder based on low-rank self-attention is designed to enhance the sequence feature representation. Then, the Decision Transformer architecture is employed to model the complex dependencies between sequence features and target behaviors. Furthermore, a weighted contrastive learning method is proposed to distinguish the preference representations under various targets. Extensive experiments on three real-world E-commerce datasets demonstrate the superiority of TFSN over state-of-the-art methods.