A Preference-driven Conjugate Denoising Method for sequential recommendation with side information

Sequential recommendation with side information aims to predict users’ preferred items based on user behavior sequences. Previous methods utilize attention mechanisms to capture user preferences from behavior sequences but often neglect individual behavioral variations, which introduce varying levels of frequency and random noise, thus compromising preference identification and integration. To address this issue, we propose a Preference-driven Conjugate Denoising Method (PCDM) for sequential recommendation with side information. The method employs a conjugate denoising transformer, consisting of a Fourier denoising module for frequency noise elimination and a variational inference module for random noise reduction, followed by a conjugate transformer that learns the user preference representations. Subsequently, it utilizes a preference-driven denoised fusion module to integrate the learned representations, aligning them with true user preferences while minimizing mixed noise interference. Experiments on four datasets, including Amazon Beauty, Sports, Toys, and Yelp, report average gains of 8.39% in Recall@10, 9.16% in Recall@20, 6.14% in NDCG@10, and 6.94% in NDCG@20 compared to the latest models.