A Collaborative Information Enhanced Mamba-Based Sequential Recommendation Model

Sequential recommendation (SR) has garnered extensive attention for its strong capacity to model evolving user preferences. While Transformer-based SR models have shown promising recommendation performance, the calculation of selfattention involves quadratic complexity, resulting in considerable computational overhead. Recent advancements in state space models (SSMs) have introduced Mamba, a novel SSM that enhances model efficiency while maintaining competitive performance compared to Transformers. Consequently, several SR models have adopted Mamba and demonstrated promising performance. However, in our research, we observe that existing Mamba-based SR models focus solely on item transitions within the current sequence, overlooking the collaborative information across sequences. This presents an opportunity to enhance recommendation performance by incorporating collaborative information cross sequences. In this paper, we propose a Collaborative Information enhanced Mamba-based Sequential Recommendation model (CIM-SR). CIM-SR proposes to capture both item transitions within the current sequence and collaborative information across sequences. Specifically, CIM-SR builds a global graph across sequences to enrich item representations with collaborative information. Additionally, CIM-SR leverages a Mamba-based sequence encoder in a sequential view to capture item transitions within the current sequence. Furthermore, CIM-SR proposes to conduct contrastive learning on item representations derived from the global graph and the sequential views to enhance item representations. Extensive experiments on three widely adopted datasets indicate that CIM-SR achieves superior performance compared to baseline models.