Enhancing E-Commerce Revenue Prediction Using Ensemble Models and Feature Engineering: A Comparative Study

This research explores the application of various machine learning models to predict e-commerce revenue using user behavior data. The study begins with an exploratory data analysis (EDA) to identify patterns and relationships among critical features, such as ‘Page Value’ and ‘Bounce Rate,’ which are crucial for predicting revenue-positive sessions. Preprocessing techniques, including data cleaning, categorical encoding, and class balancing with SMOTE, were applied to improve the dataset's quality. Several machine learning models were trained and evaluated based on their performance, including Logistic Regression, Support Vector Classifier, K-Nearest Neighbors, XGBoost, and Random Forest. Feature importance analysis was used to refine the input features, and feature reduction was applied to optimize model efficiency. Ensemble models, particularly XGBoost and Random Forest, demonstrated superior accuracy and generalization capabilities. Among the models tested, XGBoost achieved the highest performance with a test accuracy of 93.8%, making it the most effective approach for predicting e-commerce revenue outcomes.