Predicting Abnormal Events in Urban Rail Transit Systems with Multivariate Point Process

Xiaoyun Mo, Mingqian Li, Mo Li

Abstract: Abnormal events in rail systems, including train service delays and disruptions, are pains of the public transit system that have plagued urban cities for many years. The prediction of when and where an abnormal event may occur, can benefit train service providers for taking early actions to mitigate the impact or to eliminate the faults. Prior works rely on rich sources of sensor or log data that require extensive efforts in sensor deployment, data gathering and preparation. In this article, we aim at predicting abnormal events by leveraging only basic information of historical events (e.g., dates, technical causes) that can be easily obtained from existing open records. We propose a non-trivial method which categorizes event pairs based on their basic information, and then characterizes inter-event influence between event pairs via a multivariate Hawkes process. The proposed method overcomes the major hurdle of data sparsity in abnormal events, and retains its efficacy in capturing the underlying dynamics of event sequences. We conduct experiments with a real-world dataset containing Singapore’s 5-year abnormal rail events, and compare with a wide range of baseline methods. The results demonstrate the effectiveness of our method.