Research on hybrid attention sequential recommendation model based on discrete cosine transform frequency enhancement

Abstract Sequential recommendation systems aim to capture both short-term and long-term user preferences by modeling temporal patterns in user behavior. Recently, frequency-domain models have gained attention for their ability to capture global patterns efficiently. However, most existing methods rely on Fourier transforms, which are prone to Gibbs phenomenon when processing non-periodic signals, introducing high-frequency noise, and incurring substantial computational overhead due to complex-number operations. To overcome these issues, we propose Discrete Cosine transform-enhanced Hybrid Frequency-Attention Network for Sequential Recommendation (DCAN-Rec), a hybrid frequency-attention model that combines frequency information with self-attention to improve recommendation accuracy and reduce computation costs. Instead of Fourier transforms, DCAN-Rec uses discrete cosine transforms to optimize frequency-domain representations, reduce computational overhead, and mitigate high-frequency noise interference. It employs a learnable filter-based multilayer perceptron to separately extract features from high- and low-frequency components of user behavior. Experiments on seven real-world datasets show that DCAN-Rec effectively captures both global trends and local patterns in user sequences, consistently outperforming state-of-the-art methods. On average, DCAN-Rec improves performance by 16.29% over SASRec (Self-Attentive Sequential Recommendation), and also surpasses FMLPRec (Filter-Enhanced MLP for Sequential Recommendation) by 8.96%, FEARec (Frequency-Enhanced Hybrid Attention Network for Sequential Recommendation) by 3.99%, and DuoRec (Contrastive Learning-Based Model for Alleviating Representation Degeneration in Sequential Recommendation) by 3.75%, demonstrating strong generalization and stability.