Predicting biking preferences in Kigali city: A comparative study of traditional statistical models and ensemble machine learning models

This research enhanced the prediction of biking preferences in the City of Kigali, Rwanda and informed transportation management and economic policy. Specifically, it compared the performance of traditional statistical models—logistic regression, support vector machine (SVM), Naïve Bayes, and k-Nearest Neighbours (KNN)—with ensemble models including eXtreme Gradient Boosting (XGBoost), Light GBM, Random Forest, and stacking classifiers. This research used a dataset of 6386 observations incorporated weather and air quality variables and applied correlation-based and iterative model-based feature selection techniques to improve predictive accuracy. Results indicate that ensemble models, particularly XGBoost and Random Forest, outperform traditional statistical models, with an accuracy of 99 % and 98 %, respectively. Traditional statistical models underperformed, with 42 % and 82 % accuracy, in the logistic and SVM models. Ensemble models classified better biking preferences (shared, non-shared, and both categories), significantly improving precision and recall across all three groups. Feature importance indicated that day and month are critical factors in bike preference prediction, reflecting significant daily and seasonal patterns. Air quality factors (high ozone and PM2.5) and weather factors (temperature and rainfall) impacted the preferences. It is better to maintain bikes during the rainy season and rebalance bikes during high temperatures for efficient biking. To improve the air quality in the city, the government should increase car-free corridors to improve the air quality and motivate bike users to be comfortable. In a city with extreme weather, shaded bike lanes should be provided to encourage riders during the extreme weather.