Graph adaptive mix normalization in bundle recommendation: Balancing popularity bias and long-tail effect

Abstract This paper investigates the balancing problem between popularity bias and the long-tail distribution in bundle recommendation systems. Existing approaches often focus exclusively on recommending popular items or solely on recommending long-tail items. Such approaches fail to effectively balance their respective impacts, resulting in insufficient diversity in system recommendations. To address this issue, we propose the GAMNBRec model. The model introduces the graph-adaptive mixed normalization (AdaMixNorm) method, which dynamically adjusts normalization strategies based on interactive graph structures. This strategy balances recommendation performance between popular and long-tail items. It enhances the recommendation quality of long-tail items through adaptive normalization adjustments, while simultaneously preserving the quality of popular item recommendations. In addition, GAMNBRec incorporates a residual-enhanced dynamic feature fusion mechanism to prevent feature loss for long-tail items in deep networks. It also integrates a Softmax-weighted BPR contrastive loss, which dynamically adjusts the importance of negative samples and thereby improves model training effectiveness. Experiments on the NetEase, Youshu, and iFashion datasets demonstrate that GAMNBRec significantly outperforms existing state-of-the-art methods in both Recall and NDCG metrics. These results validate the effectiveness and innovation of GAMNBRec in balancing recommendations for popular items and long-tail items.