Contents lists available at ScienceDirect Journal of Rail Transport Planning & Management journal homepage: www.elsevier.com/locate/jrtpm Improvement of timetable robustness by analysis of drivers' operation based on decision trees Yasufumi Ochiai a , Yoshiki Masuma b , Norio Tomii b,∗ a Odakyu Electric Railway Co., Ltd., 1-8-3 Nishi-Shinjuku, Shinjuku, Tokyo, 160-8309, Japan b Chiba Institute of Technology, 2-17-1 Tsudanuma, Narashino, Chiba, 275-0015, Japan ARTICLEINFO Keywords: Robustness Data mining Decision tree Drivers' operation ABSTRACT In railways where trains are running densely, once there occurs a delay, even if it is small, the delay easily propagates to other trains. In order to make their timetables more robust, railway companies are making various kinds of eforts. But until now, they have not been interested in analysis of drivers’ operation, although there exists much diference in their manner of driving and the diference is closely related with robustness. Thus, it would be useful if we can know what is “good driving”, in other words, a driving which reduces a delay and what is “poor driving” meaning a driving which increases a delay. If we can know the diference between “good” and “poor” driving, we can give advice to drivers so that they can improve their driving. We have developed an algorithm to fnd the factors which diferentiate between “good” and “poor” driving based on the decision tree, which is a commonly used technique in data mining. The inputs of our algorithm are track occupation records. The algorithm receives “good” ex- amples and “poor” examples as input, then it produces a decision tree from which we can know thedominantfactorstodiferentiatebetweenthegoodexamplesandthepoorexamples.Wehave applied our algorithm to actual data and proved that the algorithm can fnd a pattern of driving which is common to poor drivers. 1. Introduction Although railways in Japan are known to be very punctual, we often encounter small delays of trains during morning rush hours in big cities especially in Tokyo. This is partly because in Tokyo area, trains are running very densely. Hence, once there occurs a delay,evenifitissmall,thedelayeasilypropagatestoothertrains(aknockondelay).Themajorcauseofsuchdelaysisanexcessof dwell times. When a dwell time of a train increases, the succeeding train cannot arrive at the station because the track is still occupied.Thus,thesucceedingtrainiscompelledtostopbeforeitarrivesatthestation.Becausesometimeisneededforthetrainto accelerate and decelerate, the train arrives at the station with a delay larger than that of its preceding train. Railwaycompanieshavebeenintensivelyexaminingcountermeasurestopreventtheincreaseofdwelltimes.Itisnowknownthat countermeasuressuchasdeploymentofsufcientnumberofstationstafonaplatform,slightrevisionoftimetables,improvementof signalling systems are efective (Yamamura et al., 2013; Yabuki et al., 2015).Butitisdesiredtofndanefective countermeasure to restore delays preferably without investing much. In this paper, we focus on drivers’ operation. From our preliminary analysis which we will show in 4.2, we found that some driverssucceedtorestoredelaysbutsomedriversareratherexpandingdelaysalthoughtheinitialsituationsaresimilar.Wesuppose https://doi.org/10.1016/j.jrtpm.2019.03.001 Received 3 April 2018; Received in revised form 4 March 2019; Accepted 5 March 2019 ∗ Corresponding author. E-mail addresses: yasufumi.ochiai@odakyu-dentetsu.co.jp (Y. Ochiai), tomii@cs.it-chiba.ac.jp (N. Tomii). Journal of Rail Transport Planning & Management xxx (xxxx) xxx–xxx 2210-9706/ © 2019 Published by Elsevier Ltd. Please cite this article as: Yasufumi Ochiai, Yoshiki Masuma and Norio Tomii, Journal of Rail Transport Planning & Management, https://doi.org/10.1016/j.jrtpm.2019.03.001