Bidirectional alignment text-embeddings with decoupled contrastive for sequential recommendation

Abstract

The key challenge in sequential recommendation is to accurately predict the next item based on historical interaction sequences by learning effective sequence representations. Existing models typically optimize sequence representations using the next ground-truth item as the supervised signal. However, this approach often results in biased interest representations and neglects the benefits of bidirectional supervision, leading to incomplete sequence representations and semantic mismatches. To address these limitations, we propose ADRec for bidirectional sequence–item Alignment text-embeddings with Decoupled contrastive learning for sequential Recommendation based only on text data. ADRec combines self-supervised and supervised signals derived from intrinsic correlations in recommendation data, to enhance semantic consistency between sequence and ground-truth item representations, improving recommendation performance. Specifically, we introduce a hybrid learning mechanism that integrates an unsupervised contrastive learning paradigm to decouple sequence and item representations and supervised contrastive learning to achieve bidirectional semantic alignment. Additionally, a dual-momentum queue mechanism is devised to expand the diversity of negative samples with limited resources, optimizing the quality of user interest representations in the text modality. Extensive experiments on six public datasets show that ADRec consistently outperforms state-of-the-art methods by learning superior sequence representations. The code is publicly available at https://github.com/pppiao/ADRec.