Predicting Customer Behavior in Autonomous Retail Application: A Classification-Based Approach

Accurately predicting system behavior is essential for optimizing performance and enabling autonomous decision-making in robotics. Similarly, forecasting customer behavior is crucial to the long-term sustainability of a business. This study introduces a novel classification-based approach for predicting the specific day of a customer's next purchase in autonomous retail systems. Utilizing a proprietary dataset comprising 2,695,572 transactions from 520,202 unique customers collected over one year from a national chain store, we transition from traditional time-series to a classification methodology. Our approach leverages advanced machine learning techniques, specifically light gradient noosting machine (LightGBM), with hyperparameters optimized through grid search and cross-validation, including learning rate, number of leaves, and max depth. This model outperforms conventional algorithms, achieving a recall of 92.36%, precision of 92.43%, and an F1 score of 92.35%. Analysis of feature importance underscores the critical role of temporal purchasing patterns and transactional attributes. This methodology enables both retail and robotic systems to adapt swiftly to behavior patterns, optimizing resource management and strategic planning.