Combining behaviour modelling and machine learning techniques to predict human mobility in major disruptions

Lead Research Organisation: University of Leeds
Department Name: Sociology & Social Policy

Abstract

Existing models for transport demand are not refined for disruptions. The work described here is intended to make use of emerging big data sources to improve the ability of transport providers to respond to changed transport requirements in these circumstances. It will establish how well existing travel demand models perform in predicting the response to major disruptions based on historical data. Candidate datasets include Oyster card data from Transport for London (TFL), mobile phone data from Telefonica (made available to academics by Transport System Catapult (TSC), etc . The disruptions to be tested will include predictable short duration events such as a sporting event or adverse weather, predictable long-term disruption such as station or line closures, as well as unpredictable events such as points failure or terrorism. Special focus will be on mathematically modelling the demand and supply variations with different influencing factors that can be used for predicting the impact of future disruptions. The importance of this work will be seen primarily in capacity planning and designing crisis-response plans. When constructing new infrastructure, it will be very useful to be able to see what parts of the old network might be stretched if any disruption occurs and plan for such scenarios, possibly by adapting the new infrastructure plans. In traditional approaches, the behaviour of people in the event of a disruption is determined using surveys - either recall surveys after a disruptive event and/or questionnaires on potential behaviour in hypothetical scenarios. Both have limitations. The recall data may not be accurate while the data on hypothetical scenarios may not be realistic. There is thus a dearth of reliable data available to study behavioural choices in stressful situations created by disruption.
Improved more robust models would be developed using insights to be obtained by mining alternative data sources such as mobile phone GPS or call detail record (CDR) data and TFL Oyster card or wi-fi log data. These data sources may provide valuable insights into the reactions of the traveling public to sudden disruptions. Other researchers have suggested there may be a large difference in the responses of Twitter users as opposed to respondents to the Household Interview Travel Survey. This may reflect the different weaknesses in each data source. Being able to collate the two sources of information will help give a better picture of what the population dynamics are in response to any given disruption. Furthermore, geo-located Twitter data could be useful when used in conjunction with smart-card (oyster) data. Cities like London have one tap smart-cards on their bus network which captures only the starting point of a journey. Combining this data with geo-location tweets may better help us identify how commuters respond to disruption when trains are out of service. The use of smart card data alone may be misleading. In order to use twitter data effectively, it might be desirable to use machine learning algorithms to tag tweets, allocating each to a categories according to the sentiment expressed. This creates the data to be used in a model of how commuters are responding to disruption. Unfortunately, statistical models and machine learning are currently ill equipped on their own to deal with human responses to disruption. This sentiment analysis is a way of gauging the general mood of commuters which will impact on how they choose to respond to the situation at hand. Choice modelling may be appropriate for this purpose. Using a measure of how upset different sets of people are during the disruption may help, in real time, decide what the best course of action might be.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
ES/S501566/1 01/10/2018 31/03/2022
2277497 Studentship ES/S501566/1 01/10/2019 30/09/2023 Mark Proudfoot