Abstract: Aiming at the problems of noise interference in multi-source interaction data, semantic inconsistencies between auxiliary behaviors and target behaviors, and the difficulty of statically allocating multi-task loss weights to adapt to users’ personalized preferences in multi-behavior recommendation scenarios, this paper proposes a multi-behavior denoising-aware recommendation model based on graph meta-networks. First, a multi-behavior-aware encoder is developed to extract high-order features from multi-type interaction data using graph convolution. Second, a selective cross-behavior contrastive learning method is designed to filter false negative samples using user node similarity, thereby achieving denoising optimization at the user representation level. Third, a hard negative sample sampling strategy based on item similarity is proposed to improve the quality of item representations. Finally, a meta-learning-based dynamic weight adjustment mechanism is implemented to adaptively adjust the multi-task loss weights. Extensive experiments and ablation studies conducted on three real-world datasets demonstrate that the proposed model outperforms various state-of-the-art recommendation methods, fully validating the effectiveness of our approach.

Key words: multi-behavior recommendation, denoising, contrastive learning, hard negative sampling, multi-task learning