Deciphering the pulse of the city: An exploration of the natural features of metro passenger flow using XAI

Abstract

Urban metro systems are integral to modern public transport, making it essential to understand the factors influencing passenger flow for effective system planning and operations. Current evaluation methods for feature importance often lack precision, creating challenges in accurately profiling influential factors. Recent advancements in explainable artificial intelligence (XAI) present opportunities to enhance feature interpretability and refine natural feature profiling frameworks for metro passenger flow. This study discusses three XAI methods, i.e., LOFO, Fast-LOFO, and SHAP, in systematically evaluating feature importance in metro systems. Utilising the metro smartcard records from Suzhou, we construct a hierarchical tagging system for natural features. Each XAI method is applied to assess feature importance across key factors like time of travel, weekday status, and points of interest, allowing for a comparative analysis of their effects on passenger flow. Our findings show that while dominant features, such as travel hour and weekday status, consistently rank as the most influential across methods, variations arise in the treatment of secondary features. Tree-based models provided stable, high-level rankings, whereas SHAP offered deeper, localised insights, highlighting how specific features influence individual predictions. These differences underscore the need for a multi-method approach to achieve a complete and context-sensitive feature profile.