MR-Rec: Mutual-representation-enhanced model for sequential recommendation

초록

Recent self-attention-based sequential recommendation models have shown limitations in capturing short-term user preferences. To address this limitation, frequency domain information has been incorporated alongside time domain modeling to enhance the representation of short-term user preferences through high-frequency signals and periodic patterns. Nevertheless, existing methods often integrate temporal and spectral representations without explicitly parameterizing their relation structures before fusion, which may limit the use of complementary sequential and frequency-aware patterns. Moreover, while high-frequency components help capture short-term interests, they can also introduce noise from temporary shifts or accidental behaviors such as gift purchases or account sharing, which hinders accurate preference modeling. Therefore, when integrating both domains, it is crucial to separately parameterize temporal and spectral compatibility structures before fusion and to filter out potential noise embedded in the high-frequency components. To address these limitations, we propose Mutual-Representation-Enhanced Model for Sequential Recommendation (MR-Rec), a mutual-representation-enhanced framework that models complementary temporal and spectral preference patterns by separately parameterizing temporal, spectral, and positional compatibility scores before integration through MR self-attention. Additionally, we introduce an adaptive noise filtering technique that identifies and removes irrelevant high-frequency signals based on their magnitude characteristics. Extensive experiments on four real-world datasets demonstrate consistent accuracy gains over strong baselines across Recall and NDCG metrics. To further assess whether attention-score fusion maintains varied recommendations, we conduct a diversity evaluation in the exploratory study.

키워드