Learnable Filter with Decoupling Fusion Method for Sequential Recommendation

The aim of sequential recommendation is to predict the next possible interaction item in a sequence by interpreting the dynamic preferences derived from users' historical interactions. However, these methodologies encounter challenges in effectively leveraging side information that could offer a more precise depiction of user preferences. Furthermore, the presence of noise in the user visit sequence, caused by accidental clicks, hampers the prediction of the subsequent item, thereby undermining the effectiveness of the recommendation. To address these issues, we propose a novel model, LFDF-SR. This model combines a Learnable Filter with Decoupling Fusion Method for Sequential Recommendation. It uses an encoder-decoder architecture to model the interactive items and their corresponding side information separately. To minimize noise, the encoder component incorporates a learnable stacked filter layer to refine item embeddings. In order to more efficiently characterize side information, we innovatively integrate the heterogeneous side information from different sources into a holistic feature representation, which are then embedded through a linear layer. Then the decoder component acquires fused embeddings through a decoupled fusion method of side information, and subsequently merges the denoised item embeddings and fused embeddings using cross-attention to generate the final representation of each item. Extensive experiments on three publicly available real-world datasets demonstrate that our model outperforms nine baseline methods in recommendation performance. These results highlight the advantages of LFDF-SR in handling noise and effectively utilizing side information, showcasing its potential to enhance sequential recommendation performance.