A Hybrid MLP and CNN Architecture for Sequential Recommendation

With the proliferation of online platforms, recommendation systems have become essential tools for providing relevant information to users. Traditional recommendation methods often rely on a single user interest, making it difficult to capture the diversity and dynamic nature of user preferences. In recent years, multi-interest sequential recommendation methods have shown promising results by leveraging temporal information in user behavior sequences. In this context, we propose a novel multi-interest sequential recommendation framework that effectively captures both long-term and short-term user interests. Specifically, we combine Multi-Layer Perceptron (MLP) and Convolutional Neural Networks (CNN) to model users’ global long-term interests and recent dynamic behaviors, respectively. Meanwhile, we introduce a multi-interest fusion mechanism to integrate these diverse interest features into a more comprehensive user representation. This framework better adapts to the evolving nature of user interests over time, thus improving both recommendation accuracy and interpretability. Experimental results demonstrate the effectiveness of the proposed model across two datasets. Further ablation studies reveal that long-term interest modeling, short-term interest extraction, and the multi-interest fusion mechanism are crucial for enhancing the model’s performance.