Bundle recommendation methods considering rating data differences for online retailers.

Area of Science:

• E-commerce
• Marketing Analytics
• Recommender Systems

Background:

• Bundling is a key strategy in e-commerce, benefiting both retailers and consumers.
• User-generated product ratings are crucial for understanding customer preferences and satisfaction.
• Data sparsity and heterogeneity pose challenges in e-commerce recommendation systems.

Purpose of the Study:

• To propose a novel bundle recommendation framework leveraging rating disparities.
• To capture nuanced user preferences and unmet demands by analyzing rating differences.
• To enhance bundle recommendation accuracy and user satisfaction in e-commerce.

Main Methods:

• A two-stage recommendation method addressing data sparsity and heterogeneity.
• Stage one: Deep Singular Value Decomposition with collaborative filtering for rating matrix completion.
• Stage two: A dual-layer graph self-attention network to model user dissatisfaction and fuse heterogeneous data.

Main Results:

• Achieved 3-6% relative improvements in Normalized Discounted Cumulative Gain (NDCG) and Recall metrics.
• Demonstrated significant increases in user satisfaction with recommended bundles.
• Validated the effectiveness of analyzing rating differences for improved recommendations.

Conclusions:

• Rating disparities offer valuable insights into user behavior and latent demands.
• The proposed two-stage model effectively enhances bundle recommendation performance.
• The framework provides a valuable tool for online retailers to improve customer experience and sales.