A survey on sequential recommendation: From classical models to LLM-powered systems

Sub-headline: Researchers from Shenzhen University and University of Queensland summarize the evolution of SR, exploring generative and long-term trends

In the era of information explosion, sequential recommendation (SR) has become essential for capturing dynamic user interests. A major technical pain point is the "interest shift"—the challenge of accurately modeling how a user's preference evolves over time from short-term impulses to long-term habits. Traditional ID-based models struggle with data sparsity and fail to leverage rich multi-modal contexts. Furthermore, processing ultra-long behavioral sequences (e.g., thousands of clicks) introduces significant computational bottlenecks and memory constraints for standard attention-based architectures.
The survey re-categorizes existing SR methods based on how they construct and utilize item properties. A key highlight is the transition from discriminative to generative paradigms, where LLMs are used to enhance the semantic understanding of user intents. The researchers provide an in-depth analysis of various backbone architectures, from Transformers to the emerging State Space Models. They also discuss how data augmentation techniques, such as contrastive learning, can mitigate the cold-start problem and improve the robustness of recommendation models in real-world scenarios.
The review concludes by identifying several promising research directions, including cloud-device collaborative recommendation, data-centric SR for high-quality training, and the development of foundation models for open-domain recommendation. This work serves as a vital guide for practitioners and researchers looking to push the boundaries of sequential decision-making in AI.