Causal discovery from train network data with background knowledge

Abstract

One delayed train could influence the punctuality of other trains in its area. Currently, the Traffic Controllers of ProRail use their own knowledge to predict the delays of the trains, where they also include the delays of other trains as a factor. ProRail want to research if it is feasible to create a decision support system, where the Traffic Controllers are aided in making predictions about delayed trains and how to intervene to minimize the disruptions. One of the first stepping stones is to create a delay prediction system, and that is what this thesis focuses on. Our goal is the exploration of causal analysis applied to delay data, and this result is included in a delay prediction model. The causal relations between trains are captured in a Structural Causal Model (SCM). Creating the SCM involves two steps: finding the causal graph, and learning the assignment functions. The causal graph is identified by applying background knowledge to the PC-algorithm, which reduces the search space. The assignment functions are learned by training multiple Neural Networks. The result of this thesis is a prediction system that includes causal relations between trains, referred to as possible train interactions, as input to predict the delays of the trains at its next time tabling point. The model performs similarly to an existing model and shows potential for improvement in further research.