Dual denoising contrastive learning with multi-interest fusion for sequential recommendation

Sequential recommendation aims to predict the next item a user is likely to interact with by learning user interest representations from historical behavior sequences. Recently, a number of studies have focused on modeling user diverse preferences through multi-interest learning, which has significantly improved recommendation performance. Despite their effectiveness, existing methods still suffer from two key limitations: (1) they mainly focus on interest clustering, while neglecting the temporal evolution of user preferences; and (2) they concentrate on modeling user interests, yet fail to address the noise inherently present in behavior sequences. To tackle these limitations, we propose a dual denoising contrastive learning with multi-interest fusion for sequential recommendation, named D2MFRec. Specifically, we first design a multi-interest aggregation module that integrates both multi-interest and single-interest representations to form a comprehensive user embedding. Then, we introduce a dual denoising module to alleviate the negative impact of noisy interactions on interest modeling. Moreover, a gated fusion mechanism is employed within the aggregation module to adaptively combine multiple sources of interest signals into a unified representation. Extensive experiments conducted on three public benchmark datasets demonstrate that D2MFRec consistently outperforms state-of-the-art baselines, validating the effectiveness of our proposed approach.