Food Delivery Time Prediction using Tree-Based Ensemble Models: A Comparative Study with Explainable Artificial Intelligence

Raihanfitri Adi Kalipaksi (Universitas Mulawarman)
Haviluddin Haviluddin (Universitas Mulawarman)
Anindita Septiarini (Universitas Mulawarman)
Joan Angelina Widians (Universitas Mulawarman)
Novianti Puspitasari (Universitas Mulawarman)

Information Technology Education Journal
Vol. 5, No. 3, August (2026)

Purpose – Being able to predict delivery time accurately is important for online food delivery services, both for operational efficiency and customer satisfaction. However, this is not an easy task. Delivery time depends on many factors that are related to each other, such as courier characteristics, delivery distance, and order-related information, and these factors interact in complex ways. This study looks at how well tree-based ensemble learning models can predict food delivery time, and also uses explainable artificial intelligence so the models can still be interpreted properly. Design – This study uses 45,593 delivery records taken from Kaggle. In this study, four tree-based ensemble models were developed, namely Random Forest, Gradient Boosting, XGBoost, and LightGBM, with each model optimized through hyperparameter tuning. The models were evaluated using repeated 5-fold cross-validation with three repetitions, and their performance was assessed using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R². Findings – LightGBM Tuned came out with the best numerical performance, showing an MAE of 5.678 minutes, RMSE of 7.204 minutes, and R² of 0.411. However, based on ANOVA and Tukey's post-hoc test, the difference in performance between the best boosting-based models was not statistically significant. SHAP analysis also showed that courier rating, delivery distance, courier age, and several interaction features were the factors that had the biggest influence on delivery time prediction. Research implications – The findings suggest that boosting-based ensemble learning models can provide moderate predictive performance while offering interpretable insights into the factors contributing to delivery time predictions. Nevertheless, the moderate R² value indicates that additional operational variables, such as traffic conditions, restaurant preparation time, and courier workload, may be required to improve practical prediction reliability. Originality/value – This study combines ensemble learning, feature engineering, statistical validation, and explainable artificial intelligence together to evaluate both the predictive performance and interpretability of models for food delivery time prediction.