Predicting Customer Purchase Intent in E-Commerce Platforms Using Session-Based Logistic Regression Analytics

Abstract

The rapid acceleration of global e-commerce has fundamentally transformed retail operations, yet digital merchants continue to confront substantial financial losses resulting from inefficient inventory planning, cart abandonment, and imprecise consumer behavior forecasting. This research develops ShopPredict, a session-based predictive analytics framework designed to forecast online shopping purchase completion by modeling user interaction dynamics including product viewing frequency, cart additions, session price averages, and unique item explorations using a Logistic Regression classification approach. Utilizing an empirical multi-category e-commerce dataset comprising millions of user session logs, the classification model was trained and evaluated across structured behavioral features. The experimental results demonstrated a high overall accuracy of 93.48%; however, rigorous metric decomposition revealed a critical accuracy paradox, wherein the model achieved a high precision of 0.94 for non-purchase sessions but exhibited a low recall of 0.16 and an F1-score of 0.25 for purchase events due to severe underlying class imbalance. System usability and functional architecture evaluations yielded positive user acceptance scores averaging above 4.0 on a 5-point Likert scale. This study underscores the utility of session-based linear probability modeling while demonstrating that high nominal accuracy in e-commerce prediction can be misleading without rigorous class-rebalancing strategies. Future research must incorporate advanced ensemble architectures, SMOTE oversampling, and dynamic session features to optimize purchase intent detection in real-time retail environments