Comparative Evaluation of Resampling Techniques on Machine Learning for Predicting Customer Purchasing Behavior

In the era of subscription-based services, customer purchasing behavior presents a significant challenge for organizations because customer retention is essential to long-term success. Imbalanced datasets, where the majority class significantly outweighs the minority class, often cause models to favor the dominant class, leading to biased predictions. This study investigates customer purchasing behavior, focusing specifically on the issue of class imbalance. It offers a comparative analysis of various classification techniques and includes widely used resampling methods such as the Synthetic Minority Over-sampling Technique (SMOTE), Edited Nearest Neighbors, Tomek Links, and several hybrid sampling approaches. Our research findings demonstrate the competitive effectiveness of using the Random Forest algorithm with undersampling methods such as TomekLinks and Edited Nearest Neighbors (ENN), which have shown the highest F1 scores. However, the hybridization of oversampling and undersampling yields unsatisfactory results, below our expectations. This suggests that the dataset used in this study exhibits a high degree of overlap between the two classes. Consequently, undersampling proves to be an effective method for mitigating the negative impact of class overlap. This research makes a substantial contribution to the field of predicting customer purchasing behavior.