Multi-Class Machine Learning Models for Predicting Order Cancellations and Returns in E-Commerce

E-commerce became a central component of sustainable growth strategies for businesses, driven by evolving consumer behavior and accelerating digitalization.In this study, multi-class machine learning models were developed for predicting order cancellations and returns in the e-commerce sector. A real-world dataset containing $\mathbf{1 0, 0 0 0}$ order records collected from a client of Inveon, covering the period from 2020 to 2024, has been used.The performance of five classification algorithms, including Logistic Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Deep Neural Network (DNN), was evaluated and compared. Four feature selection techniques were applied to enhance predictive performance, including minimum Redundancy Maximum Relevance (mRMR), Relief-F, FClassification, and a hybrid approach. Model generation has been carried out using $\mathbf{1 0}$-fold cross-validation and $\mathbf{3}$-fold grid search for hyperparameter tuning. RF- and XGBoost-based models achieved the highest performance, with F1-scores reaching up to 0,96. Regarding class-specific prediction capability, RF-, XGBoost-, SVM-, and DNN-based models yielded better results in cancellation prediction, while LR-based models performed more effectively in return prediction. Among the feature selection techniques, Relief-F and the hybrid method made the most notable contributions to model performance, especially in RF- and XGBoost-based models.