A Sequential Recommendation Model Based on Convolutional Neural Networks and State Space Models

Abstract

Sequential recommendation aims to predict users' future preferences by analyzing their historical behavior sequences. In recent years, deep learning techniques have been widely applied to sequential recommendation tasks, achieving remarkable improvements in recommendation accuracy. However, existing methods often suffer from structural imbalance in modeling the rapidly changing short-term interests and the relatively stable long-term preferences of users. Moreover, these methods face limitations such as high computational costs and low inference efficiency when dealing with long behavior sequences. To address these challenges, this paper proposes a novel sequential recommendation model named CNN-Mamba, which integrates Convolutional Neural Networks (CNN) with State Space Models (SSM). Specifically, the CNN component leverages local receptive fields to efficiently extract short-term interest features from recent user interactions, thereby enhancing the model’s ability to capture local behavior patterns. Meanwhile, the SSM component is introduced as a long-term interest modeling module to capture global dependencies within long sequences. Furthermore, an adaptive fusion layer is designed to dynamically integrate the short- and long-term modeling outputs, thereby improving the model's generalization ability. In addition, the implicit recurrence mechanism in the state space model effectively reduces computational complexity and enhances the efficiency of long-sequence modeling. Experimental results on three real-world datasets demonstrate that the proposed CNN-Mamba model outperforms state-of-the-art baselines in both recommendation accuracy and inference efficiency, validating its effectiveness and practicality.