This research proposes an interaction-aware framework to address three key gaps in metro ridership studies: single-factor dominance, neglect of extreme demand scenarios, and lack of circleline-based spatial differentiation. Using Guangzhou’s metro system, the framework integratesensemble machine learning with SHAP interaction analysis to examine how built environment,network topology, and intermodal accessibility jointly influence ridership. The results show thatnetwork topology is the dominant determinant, and key single factors—especially betweennesscentrality—remain central within top interaction pairs. Under extreme demand, ridership isprimarily governed by structural network characteristics rather than temporal variation. Clearspatial heterogeneity is observed: downtown ridership is driven by topology–topology interactions, while peripheral ridership depends more on cross-dimensional interactions, particularlyaccessibility to the metropolitan center. Interaction effects account for a substantial share ofridership formation, with several variables exerting influence mainly through interactions.