Modeling women cyclists' perceived security: A comparison of machine learning techniques

Incorporating human behavioral factors into travel demand analysis is increasingly critical in transportation planning. Perceived security is a key behavioral factor influencing urban travelers’ choices, particularly for active transportation modes like cycling and vulnerable road users like women. While prior studies have explored certain aspects of perceived security —mainly for pedestrians—the application of Machine Learning (ML) models to predict cyclists’ perceived security remains a relatively developing research area. This study estimates women cyclists’ perceived security by leveraging environmental features derived from data collected through a virtual reality (VR)-based bicycle experimental system. An inferential approach was employed, combining Automated ML (AutoML) tools, consecutive Grid Search (GS) and Random Search (RS) hyperparameter optimization, and various ensemble techniques to provide insights into the comparative performance of available ML models. Although the model selection in the initial stage could be inferred from travel behavior modeling literature, we included AutoMLs in the initial stage to gain insight into the comparative performance of more diverse (68) ML models. This systematic approach enabled the selection of the most promising models, which were refined through subsequent stages. The final model, an ensemble configuration labeled as DNN-stack2, integrates a Deep Neural Network (DNN) meta-learner with three base learners: light gradient-boosting machine (LGBM), k-nearest neighbors (KN), and extra trees (ET). Achieving an accuracy of 84.07 % and leveraging feature importance analysis to enhance its interpretability, the DNN-stack2 represents a marked improvement over previous models, underscoring the potential of advanced ML techniques for modeling perceived security among women cyclists.