Long- and short-term preferences modeling based on dual-frequency self-attention network for sequential recommendation

Sequential recommendation aims to analyze users’ interaction sequences to capture their sustained long-term preferences and dynamically changing short-term preferences for the next item recommendation. Recent studies have shifted their focus to the frequency domain to further mine users’ complex historical interaction behaviors. However, most existing frequency-based methods cannot explicitly distinguish the low-frequency information associated with long-term preferences from the high-frequency information associated with short-term preferences in user sequences. Consequently, they are unable to accurately model these preferences, thereby limiting the performance of the models. To this end, we propose a novel yet simple model based on Dual-Frequency Self-Attention Network (DFSNet) for sequential recommendation. DFSNet comprises low- and high-frequency self-attention modules that separately extract the corresponding components from user sequences to model long- and short-term preferences. Additionally, considering the limited frequency information available within sequences, we introduce contrastive learning to generate self-supervised signals from the preference representations produced by DFSNet. This approach further strengthens the modeling of both long-term and short-term preferences without disrupting the sequence structure, thereby positively impacting the recommendation performance. Extensive experiments on four public datasets indicate that DFSNet outperforms strong baselines while balancing accuracy and efficiency, confirming its effectiveness.