Resilience of public transport in the face of disruptions: Insights from explainable machine learning

Disruptions in public transport (PT) can have a major impact on passenger activities and on the attractiveness of the service, particularly when they are not absorbed by the network as a whole. The present study aims to detect the presence of disruption and assess the contribution of existing alternative bus or tramway stops to the resilience of the PT network, using explainable machine learning techniques. The detection task is formulated as a supervised classification problem performed using Random Forest (RF) for 39 different subway stations, using Automatic Fare Collection (AFC) data and Service Disruption logs (SD-logs). Furthermore, the SHapley Additive exPlanation (SHAP) interpretation method is implemented to retrieve the magnitude and the direction of each alternative stop's contribution to PT resilience. Results show that the proposed modeling framework has high prediction performance, can minimize false alarm rates, and can foresee the occurrence of disruptions 5 min before their registered beginning in SD-logs. Findings also indicate where demand is reallocated, resulting in 5 different resilience clusters for subway stations. Density and connectivity emerge as two major attributes of resilience that have a central role in the design of disruption management (tactical) and development (strategical) plans. The proposed approach has been applied to the PT network of Lyon (France) and is replicable by adapting the hyperparameters to the observed use in other PT networks.