Sequential Recommendation Model Based on Frequency-Aware Multi-Scale Fusion and Hierarchical Contrastive Learning

Self-attention mechanisms have achieved significant progress in sequential recommendation. However, existing studies reveal that self-attention-based methods tend to suffer from representation over-smoothing, which undermines the recommendation performance of personalized recommendation. To address this issue, this paper proposes a sequential Recommendation model based on Frequency-Aware multi-scale fusion and hierarchical Contrastive Learning (FACLRec) . FACLRec adopts a dual-branch structure operating in both the frequency and time domains. The frequency branch combines Fourier-based low-pass filtering and depthwise convolution to capture global trends and local high-frequency fine-grained features, while the time-domain branch integrates a central-node mechanism with convolution-enhanced multi-scale attention to capture multi-granularity temporal dependencies. The representations learned from the two branches are adaptively fused with learnable weights to enhance feature diversity. Furthermore, a hierarchical contrastive learning strategy is introduced to increase the discriminability across encoding layers, effectively alleviating the over-smoothing problem. Experimental results on five public recommendation datasets demonstrate that FACLRec consistently outperforms six representative baseline methods across multiple evaluation metrics, achieving up to 8.01% improvement in Hit Rate, and showing clear advantages in handling long sequences and sparse data scenarios, confirming its effectiveness.