A method for B2B customer classification based on theutility-weighted temporal profile in E-commerce systems

Abstract

Purpose: The study addresses the problem of predictive content personalization in business-to-business e-commerce systems, where classical recommendation and segmentation approaches demonstrate limited effectiveness due to the financial heterogeneity, cyclic purchasing behavior, and long-term nature of customer interactions. The purpose of the research is to develop a method for classifying business-to-business customers based on a utility-weighted temporal customer profile, which makes it possible to generate more relevant and timely predictive offers. Objectives: The proposed method is designed as the third stage of an analytical pipeline that extends utility pattern mining and temporal analysis of stable purchasing cycles. Methods: At the first stage, financially significant product patterns are identified using the Utility Pattern Growth approach. At the second stage, the temporal stability of repeated purchases is assessed through inter-purchase time intervals and the coefficient of variation. At the third stage, the obtained pattern-level results are aggregated into an integrated customer profile that combines the average purchasing cycle, cycle stability, and average transaction value. These features are normalized and used for customer clustering, after which customer classes are integrated into the mechanism for determining an individual communication trigger window. Results: The experimental validation was performed on historical transactional and behavioral data from the business-to-business e-commerce platform “Baza Vzuttya” for the period from two thousand twenty-two to two thousand twenty-five. The comparative experiment included four recommendation scenarios: a baseline frequency-based approach, a utility-weighted approach, a utility-temporal approach, and an integrated approach with customer classification. The results show a consistent increase in recommendation conversion from three point nine one percent in the baseline scenario to fifteen point nine one percent in the integrated scenario. At the same time, the number of generated recommendations decreased from nine hundred and twenty to two hundred and twenty, which indicates a reduction in irrelevant communications. The integrated scenario achieved the highest average revenue per recommendation, amounting to six thousand nine hundred and nine point zero nine hryvnias. Conclusions: The obtained results confirm that the proposed method improves the accuracy, timeliness, and economic validity of personalized predictive offers in business-to-business e-commerce systems.