基于自适应探索和残差融合的会话推荐方法

his paper proposes an Adaptive Exploration and Residual Fusion (AERF) method aimed at addressing the long-tail distribution problem in session-based recommendation systems. The AERF model constructs a global item graph and employs an adaptive random walk mechanism to generate multiple high-quality walk sequences for the target item from a global perspective, adaptively selecting contextual information relevant to the target item to effectively enhance the representation of long-tail items. Additionally, the model introduces a residual fusion mechanism, where the branch predictions trained for different item categories are fused with the main branch predictions as residuals, effectively improving the recommendation performance for long-tail items. Experimental results show that, compared to existing state-of-the-art methods, the AERF model achieves significant performance improvements on multiple real-world datasets, while ablation studies validate the effectiveness of the adaptive exploration mechanism and residual fusion mechanism. This method enhances the diversity and fairness of session-based recommendation systems by significantly improving the recommendation performance for long-tail items without compromising the prediction accuracy of head items.