Multi-behavior recommendation based on intent learning

Users often exhibit different intents when interacting with recommender systems, guiding their engagement across various behavior categories like clicks, ratings, and purchases. However, most current approaches overlook this diversity of user behaviors, making it difficult to capture the varied structural linkages occurring across multiple interaction types. Additionally, prior multi-behavior recommendation research frequently neglects modeling the underlying intents motivating different activities. Consequently, the potential of leveraging behavioral data to enhance recommendation performance for target outcomes remains underutilized. Exploring behavior intent is critical for recommender systems, but poses significant challenges due to three key factors: (1) capturing the diverse intents behind multiple interaction behaviors, (2) modeling interdependencies among various user-item interactions, and (3) integrating multi-behavior signals with heterogeneous user behavior collaboration characteristics. To address these difficulties, we propose a novel model called Multi-Behavior Knowledge Graph Intent Network (MBKGIN). MBKGIN utilizes a knowledge graph to understand the intents behind behaviors, overcoming limitations of previous methods. Specifically, MBKGIN constructs multi-behavior dependencies using a multi-head attention mechanism and incorporates intent information from the knowledge graph. Experiments on real-world datasets demonstrate MBKGIN’s effective utilization of multi-behavior data.