Knowledge-constrained interest-aware multi-behavior recommendation with behavior pattern identification

Recommender systems aim to accurately capture user preferences based on interacted items. Conventional recommender systems mainly rely on the singular-type behavior of users, which may limit their ability to handle practical scenarios (e.g., E-commerce). In contrast, multi-type behavior recommendation (MBR) exploits auxiliary types of behaviors (e.g., view, cart), as well as the target behavior (e.g., buy), and has proven to be an effective way to identify user preferences from various perspectives. Existing MBR methods assume that all auxiliary behaviors of a user have a positive relevance with the target behavior. However, users may not interact with items using all available behaviors, but the degree of relatedness is not explicitly taken into account. To address the issue, we propose a Knowledge-constrained Interest-aware Framework with Behavior Pattern Identification (KIPI). The proposed model identifies user-specific behavior patterns by introducing pair-wise dependency modeling to explicitly reflect the fine-grained relatedness between behavior pairs. Additionally, we enhance item representations by leveraging both instance-view knowledge graph (KG) and ontology-view KG, which provides broader concept information of items. Moreover, we design a concept-constrained Bayesian Personalized Ranking loss to reflect a user's general interest. Extensive studies on the real-world datasets demonstrate that our model outperforms state-of-the-art baselines.