A comparative study of predicting travel mode choice of school children using explainable machine learning techniques

Abstract

Prediction of mode choice of school children is an important research topic for transportation planning. Traditionally, mode choice studies of school children are conducted using statistical or simple machine learning techniques. Though statistical techniques provide a good basis for theoretical learning and interpretability, they are mostly based on unrealistic assumptions which might lead to biased predictions. Alternatively, machine learning approaches do not provide any theoretical basis, with poor interpretability and do not provide any insights about factors affecting behavioral aspects. To fill this gap, this research proposes explainable machine learning approaches to comprehend the mode choice prediction of school children in Sahiwal, Pakistan. Data was collected from different schools in Sahiwal district through questionnaire survey and 1,498 completed responses were collected for further analysis. Different explainable machine learning techniques (such as Logistic Regression, Decision Tree, Random Forest, k -Nearest Neighbors, and Light Gradient Boosting) were developed to model the mode choice of school children. Results showed that the Random Forest outperformed as compared to other models. In order to avoid Blackbox criticism of machine learning models and improve their interpretability, variable importance and SHAP dependency analysis were also performed. The results showed that predictors such as travel cost, monthly household income, distance to school, class grade and number of family members were significantly influencing mode choice of school children. These findings can be better used for effective modeling and planning of mode choice preferences of school children.