Analyzing Network-Wide Patterns of Rail Transit Delays Using Bayesian Network Learning

Rail transit delays are generally discussed in terms of on-time performance or problems at individual stops. Such stop-scale approaches ignore the fact that delays are also caused and perpetuated by network-wide factors (e.g., bottlenecks caused by shared tracks by multiple transit lines). The objective of this paper is to develop a network model and metrics that can quantify the delay dependencies between transit network stops, and identify local sources of network-wide issues. For this purpose, Bayesian network learning (at the intersection of machine learning and network science) was utilized. Based on the calculated Bayesian networks (BNs), network metrics (inducer and susceptible) were formulated to quantify the network-wide impacts of the delays experienced at the stops. To implement the proposed framework, the delays at Long Island Rail Road (LIRR) were gathered through a crowdsourced real-time transit information app called onTime. The developed BN model was tested through cross-validation, yielded promising accuracy results, successfully identified the problematic stops based on LIRR reports, and provided further insights on network impacts. The BN model and the developed metrics were further tested using a natural experiment, i.e., a before and after study focusing on a recently completed track expansion project at LIRR. The findings imply that BN learning can successfully identify the network dependencies and indicate the rail links/corridors that are the best candidate for subsequent improvement investments. Overall, the developed metrics can quantify the delay dependencies between stops and they can be used by policy makers and practitioners for investment and improvement decisions.