Bundle Detection based on Graph Convolutional Network Considering Categorical Hierarchical Relationship

Keigo Sakurai, Ren Togo, Takahiro Ogawa, Miki Haseyama

Abstract: Item bundling is a promising sales strategy for e-commerce services, and researchers have been actively investigating techniques to create bundles that can enhance user satisfaction. A fundamental approach to creating a bundle involves detecting a set of highly relevant items as a bundle that users interact with during a single session. Previous studies have addressed the bundle detection task using non-personalized data mining methods, focusing on the observation of co-occurring items, however, these methods cannot capture the relationships and influences among items due to the discrete processing of items. In this paper, we propose a novel bundle detection method using the Graph Convolutional Network (GCN) to capture the hierarchical relationships among item categories. By utilizing GCN, our method can detect bundles that exhibit infrequent yet significant semantic and potential connections. We validate the effectiveness of our approach through comprehensive quantitative and qualitative evaluations using multiple real-world datasets.