User-Aware Denoising Sequential Model for Multi-Behavior Recommendation

Multi-behavior recommendation models address the limitations of traditional models that rely solely on sparse target behaviors by leveraging auxiliary behaviors. However, the existing multi-behavior recommendation models suffer from suboptimal performance due to two overlooked issues: behaviorlevel noise arising from inconsistent user intentions, and featurelevel noise stemming from varying user focus across item features. Previous studies have attempted to mitigate behaviorlevel noise using techniques such as attention mechanisms and self-supervised learning, but have overlooked the interference of feature-level noise. To address the above issues, we propose a User-Aware Denoising Sequential Model (named UADEN), which simultaneously denoises at both behavior and feature levels. UADEN consists of two core components: (1) A global multi-behavior feature extractor that models global user-item interactions on behavior-specific graphs using LightGCN, and dynamically fuses them via external attention to suppress behaviorlevel noise; (2) A user-aware personalized feature filter that adaptively filters irrelevant item features conditioned on user preferences, mitigating fine-grained noise at the feature level. Extensive experiments on three real-world datasets demonstrate the superiority of UADEN over 17 state-of-the-art baselines.