Explaining deep learning-based activity schedule models using SHapley Additive exPlanations

Artificial neural networks are often criticized for their black box nature in travel behavior literature. The lack of understanding of variable influence induces little confidence in model predictions, significantly affecting their practical utility. This study aims to address this issue by employing SHapley Additive exPlanations to understand the influence of different variables in a deep learning-based activity schedule model. The activity schedule is represented as a time series which enables the study of temporal variations in the influence of each variable at much finer resolutions compared to earlier approaches. The findings reveal that variables such as the day-of-week, month of the year, and social participation wield significant influence over the activity schedule, while household structure and urban class also exert noticeable impacts. This proposed methodology enhances our understanding of variable influences at different times of the day, instilling confidence in the deep learning model's results, advancing its practical application.