<p>Session-based recommendation (SBR) plays a critical role in practical scenarios such as e-commerce product suggestion and music playlist generation, where user identity is often unknown, and interactions are limited to short-term sessions. However, due to the limited short-term interaction history of users, SBR faces the challenge of data sparsity. Additionally, although existing techniques have improved model performance by introducing reverse positional information in position-aware attention mechanisms, this often compromises the sequential nature of session data. Furthermore, noise is typically introduced during the construction of session graphs and feature extraction, further deteriorating model performance. To address these challenges, this paper proposes a novel SBR technique based on the MIF-SSGCT framework. This method learns session embeddings by combining the information of three graph structure types: the global interaction graph, the local interest graph, and the behavior sequence graph. The co-training mechanism is employed to integrate multiple representations, and contrastive learning is used to mitigate the data sparsity problem. Moreover, by leveraging the behavior sequence graph that incorporates temporal information, the impact of reverse positional information on sequence integrity is effectively mitigated. Experimental results on three benchmark datasets demonstrate that this method significantly outperforms existing approaches.</p>