Filter-Enhanced Multi-interest Network for Sequential Recommendation

As the user base and the diversity of interests increase, extracting a single user representation vector has gradually become a bottleneck in modeling multiple user interests. Recent works have shown that capturing the multi-interest representation of users improves recommendation accuracy. However, user behavior sequences exist noise due to user’s mistakes or other factors. The noise and interest-irrelevant items vary for user diverse interests, and lead to inefficient multi-interest extraction and suboptimal performance. To this end, we propose a novel Filter-Enhanced Multi-Interest network for sequential Recommendation, named FEMIRec. First, we devise interest-oriented adaptive filters to mitigate the influence of noise and interest-irrelevant items for each user interest. Then, the multi-interest extractor leverages the denoised sequence representations to generate more effective multi-interest embeddings. To supervise the attenuation of interest-specific noise, we design an interest-specific noise-attenuation loss as a supervision signal, which utilizes the similarity weight calculated from the target denoised sequence and the target item as a soft label. Extensive experiments on three real-world datasets demonstrate the effectiveness FEMIRec.