Predictive Web System Based on Machine Learning to Optimize Cross-Selling: Case of a Chemical and Reagent Distribution Company

Serna-Ventura C.A.; Zavala-Zapata A.L.; Carrera-Salas E.A. · 2026 7th International Conference on Artificial Intelligence, Robotics, and Control, AIRC 2026 (2026) · DOI: 10.1109/airc69745.2026.11631324

Resumen

Small and medium-sized enterprises (SMEs) face obstacles when incorporating data-driven strategies, which restricts their ability to execute effective cross-selling and reduces their profit margins. The challenge lies in the absence of adequate tools for information processing and informed decision-making, particularly in Business-to-Business contexts where the costs of acquiring new customers are considerably higher than retention strategies. The proposed solution involves developing a web platform that integrates a predictive model using machine learning techniques, specifically utilizing the XGBoost algorithm. The model was trained with approximately 11,000 historical transaction records and 21 attributes, employing an 80% distribution for training and 20% for evaluation. The system enables predictive analysis of the probability that specific customers will purchase products belonging to particular categories. The obtained results demonstrate the model's effectiveness. During the validation phase, the XGBoost model achieved an ROC-AUC of 93.0%, indicating strong discrimination capability. When evaluated on an external test dataset of 300 additional records, the model obtained the following performance metrics: Precision 62.6%, Recall 73.4%, Accuracy 79.6%, F1-Score 67.6%, and ROC-AUC 85.6%. These indicators demonstrate the model's strong capability to distinguish between customers with high and low probability of making additional purchases.

Instituciones: Universidad Peruana de Ciencias Aplicadas