Food Delivery Time Prediction in Indian Cities Using Machine Learning Models
Ananya Garg, Mohmmad Ayaan, Swara Parekh, Vikranth Udandarao
TL;DR
This work tackles food delivery time prediction in Indian cities by integrating real-time contextual data (traffic, weather, events) and geospatial information into ML models. A comprehensive evaluation across Linear Regression, Decision Trees, Random Forest, XGBoost, and LightGBM identified LightGBM as the top performer with $R^2=0.76$ and $MSE=20.59$. Ablation and significance analyses demonstrate the critical role of real-time and geospatial features in prediction accuracy. The study provides practical guidance for dynamic routing and ETA estimation, and makes the reproducible code publicly available for further research.
Abstract
Accurate prediction of food delivery times significantly impacts customer satisfaction, operational efficiency, and profitability in food delivery services. However, existing studies primarily utilize static historical data and often overlook dynamic, real-time contextual factors crucial for precise prediction, particularly in densely populated Indian cities. This research addresses these gaps by integrating real-time contextual variables such as traffic density, weather conditions, local events, and geospatial data (restaurant and delivery location coordinates) into predictive models. We systematically compare various machine learning algorithms, including Linear Regression, Decision Trees, Bagging, Random Forest, XGBoost, and LightGBM, on a comprehensive food delivery dataset specific to Indian urban contexts. Rigorous data preprocessing and feature selection significantly enhanced model performance. Experimental results demonstrate that the LightGBM model achieves superior predictive accuracy, with an R2 score of 0.76 and Mean Squared Error (MSE) of 20.59, outperforming traditional baseline approaches. Our study thus provides actionable insights for improving logistics strategies in complex urban environments. The complete methodology and code are publicly available for reproducibility and further research.