Dynamic and Interpretable Hazard-Based Models of Traffic Incident Durations

Kieran Kalair,Colm Connaughton

doi:10.3389/ffutr.2021.669015

Kieran Kalair, Colm Connaughton

Open Access

https://doi.org/10.3389/ffutr.2021.669015

Copy DOI

Abstract

Understanding and predicting the duration or “return-to-normal” time of traffic incidents is important for system-level management and optimization of road transportation networks. Increasing real-time availability of multiple data sources characterizing the state of urban traffic networks, together with advances in machine learning offer the opportunity for new and improved approaches to this problem that go beyond static statistical analyses of incident duration. In this paper we consider two such improvements: dynamic update of incident duration predictions as new information about incidents becomes available and automated interpretation of the factors responsible for these predictions. For our use case, we take one year of incident data and traffic state time-series data from the M25 motorway in London. We use it to train models that predict the probability distribution of incident durations, utilizing both time-invariant and time-varying features of the data. The latter allow predictions to be updated as an incident progresses, and more information becomes available. For dynamic predictions, time-series features are fed into the Match-Net algorithm, a temporal convolutional hitting-time network, recently developed for dynamical survival analysis in clinical applications. The predictions are benchmarked against static regression models for survival analysis and against an established dynamic technique known as landmarking and found to perform favourably by several standard comparison measures. To provide interpretability, we utilize the concept of Shapley values recently developed in the domain of interpretable artificial intelligence to rank the features most relevant to the model predictions at different time horizons. For example, the time of day is always a significantly influential time-invariant feature, whereas the time-series features strongly influence predictions at 5 and 60-min horizons. Although we focus here on traffic incidents, the methodology we describe can be applied to many survival analysis problems where time-series data is to be combined with time-invariant features.

Highlights

Managing and reducing congestion on roads is a fundamental challenge faced across the world
We note that elastic net regularization is applied to all deep learning methods, and the optimal Cox and Accelerated Failure Time (AFT) models are selected though inspection of sample-size adjusted Akaike information criterion (AIC) to avoid over-fitting
The only difference here is we are computing the concordance index (C-index) at a given prediction time and horizon rather than over the entire dataset, and this is the definition given in Lee et al (2020)

Summary

Introduction

Managing and reducing congestion on roads is a fundamental challenge faced across the world. While a significant component of the United Kingdom transport infrastructure, congestion remains a major problem on the network, with the cited report suggesting 75% of businesses consider tackling congestion on the SRN is important or critical to their business. A traffic incident is considered to have four different time-phases: the time taken to detect and report an incident, the time to dispatch an operator to the scene, the travel time of the operator to the scene, and the time to clear an incident. Such a framework is described in Li et al (2018). The idea to use such a speed profile is considered in Hojati et al (2014), where they define the “total incident duration” to be the time from incident start until the speed has recovered to the profile

Methods

Results

Conclusion