Presentation Information
[4Yin-B-40]A Study on a Scalable Method for Customer Value Estimation
〇Rikuto Yoshizawa1, Daichi Ojiro1, Shigenori Matsumoto1 (1. Hitachi)
Keywords:
Values,LLM,Purchase history
In today’s marketplace, where customer values are increasingly diverse, a key challenge in marketing is to design effective appeal strategies that appropriately capture product value. As a method for analyzing product value, the Means–End Chain (MEC)—which links product attributes to values via benefits—has long been used; however, the high cost of collecting interview-based data remains a major bottleneck. In this study, we propose a method to infer customers’ underlying values from purchase histories, which can be collected at much larger scale. Because purchase situations can bias inferred values depending on the characteristics of the products sold, our method removes such bias using a hierarchical Bayesian model. In addition, we adopt Schwartz’s theory of basic human values and refine the estimated values by leveraging their circumplex (circular) structure. For evaluation, we used purchase histories generated by virtual customers endowed with specific values, and measured the cosine similarity between the value vectors estimated by the proposed method and the ground truth. Results show that the proposed method improves cosine similarity by 0.37 compared with a baseline approach that feeds the entire purchase history to an LLM in a single prompt.