When feature encoder meets diffusion model for sequential recommendations

Abstract

The key to sequential recommendations is to learn accurate item embeddings based on users' historical interactions. Existing methods rely on either fixed vectors or distributions to represent items. Though these methods are effective, we argue there are two limitations. a) User preferences for items can arise from multiple aspects. Fixed vectors are insufficient to capture the multifaceted features of items and the user's novel intentions. b) The conventional constraints of distribution representations during modeling process lead to the loss of some flexibility. To address these limitations, we propose the Hybrid Representation model for Sequential Recommendation (HR4SR), which utilizes both fixed vectors and distribution representations to model the features in interaction sequences. Specifically, we propose the use of diffusion techniques to introduce distribution vectors by gradually adding noise, which can compensate for the inadequacies of fixed vectors. In addition, we remove the traditional constraints of diffusion techniques, allowing distribution vectors to capture the fine-grained features of items more flexibly. Finally, we utilize a combination of fixed vectors and distributed vectors to form the item embeddings, where distributed vectors are means to compensate for the fixed vectors' inability to capture finer-grained user preferences. Experiments on three datasets show that HR4SR significantly outperforms strong baselines. The code is released at https://github.com/xiaocilu1999/HR4SR.