Discovering Cyclists' Street Visual Preferences From Shared Bike Trajectories and Street View Images Using XAI-Enhanced Inverse Reinforcement Learning

Abstract

Cycling has gained global popularity due to its health benefits and positive impacts on urban areas. While prior work has examined how urban built environments influence cyclists’ behaviors, cyclists’ route-choice mechanisms remain insufficiently understood largely due to overlooking the complex and continuous nature of cycling. To address this gap, we introduce a data-driven framework that integrates Maximum Entropy Deep Inverse Reinforcement Learning (MEDIRL) with explainable artificial intelligence (XAI). This framework is designed to quantify and interpret cyclists’ street visual preferences, which are their preferences for the street-level built environment during cycling. Our approach uses MEDIRL to approximate reward functions that reflect cyclists’ street visual preferences by integrating dockless bike-sharing (DBS) trajectories with street view images (SVIs). Subsequently, we apply XAI techniques to reveal the complex effects of street visual elements on those quantified street visual preferences. We validate our framework in the Bantian Subdistrict, Shenzhen. The results demonstrate that the quantified street visual preferences significantly outperform shortest-path models and other baselines in explaining cyclists’ route decision-making. Besides, cyclists exhibit strong preferences towards elements associated with perceived safety, street enclosure, and cycling comfort. Moreover, our detailed analysis discovers nonlinear relationships among these street visual elements, providing a scalable, data-driven foundation for understanding cyclists’ behaviors and guiding streetscape design.