Bundle recommendation methods considering rating data differences for online retailers

Bundling has emerged as a pivotal marketing strategy for online retailers, offering mutual benefits to both merchants and consumers in the rapidly expanding e-commerce landscape. Among various types of user behavior data, user-generated product ratings serve as a critical indicator of individual preferences and satisfaction levels. This research proposes a novel bundle recommendation framework that leverages rating disparities to capture nuanced user preferences and unmet demands. To address the challenges of data sparsity and heterogeneity, we develop a two-stage recommendation method. In the first stage, we enhance the completion of sparse rating matrices by integrating collaborative filtering with deep singular value decomposition. A modified cosine similarity function is introduced, incorporating a rating correction coefficient and an item popularity coefficient to improve similarity estimation. In the second stage, we exploit insights from low-rated items to model user dissatisfaction and latent demands. A dual-layer graph self-attention network is constructed to fuse heterogeneous data, refine inter-item relational representations, and enhance bundle recommendation accuracy. Extensive experiments conducted on benchmark Amazon datasets demonstrate the effectiveness of our approach, achieving 3–6% relative improvements in NDCG and Recall metrics compared to state-of-the-art baselines. Moreover, user satisfaction with the recommended bundles also increased significantly. These results highlight the value of rating differences in understanding user behavior and validate the efficacy of our two-stage model in improving bundle recommendation performance for online retailers.