Sequential recommender systems: A methodological taxonomy and research frontiers

In the era of information overload, sequential recommender systems have emerged as pivotal tools for modeling user preferences through dynamic behavioral pattern mining. These systems transcend conventional recommendation paradigms by explicitly modeling temporal dependencies in user–item interactions, preference evolution, and contextual dynamics. This study presents a methodologically structured taxonomy of sequential recommender systems through four analytical dimensions: (1) Sequential Modeling, which includes methods ranging from statistical techniques to deep learning architectures to understand user behavior patterns; (2) Temporal Dynamics Modeling, which involves time-aware collaborative filtering and deep temporal modeling; (3) Network-Enhanced Modeling, which leverages graph neural networks, heterogeneous graphs, dynamic graphs, and hypergraphs to explore structural dependencies; and (4) Robust Representation Learning, which encompasses contrastive mechanisms and techniques driven by large language models (LLMs). These algorithms focus on different aspects of sequential recommendation, including but not limited to capturing dynamic interests, modeling long- and short-term preferences, and addressing issues such as data sparsity, noise, and bias, which affect the performance and user experience of recommender systems in practical applications. Furthermore, we summarize and discuss promising future research directions to provide theoretical and methodological insights. The constructed taxonomy not only organizes existing methodological innovations, but also reveals fundamental limitations in current evaluation protocols, providing a roadmap for advancing both theoretical foundations and practical applications in this domain.