Dynamic and interpretable hazard-based models of traffic incident durations

TL;DR

This paper introduces dynamic, interpretable hazard-based models for predicting traffic incident durations using real-time data and machine learning, improving prediction accuracy and interpretability over static methods.

Contribution

It develops a novel dynamic prediction approach with real-time updates and interpretable feature importance for traffic incident durations.

Findings

Match-Net outperforms static models and landmarking in prediction accuracy.

Time-invariant features like time of day are consistently influential.

Time-series features significantly impact short- and long-term predictions.

Abstract

Understanding and predicting the duration or "return-to-normal" time of traffic incidents is important for system-level management and optimisation of road transportation networks. Increasing real-time availability of multiple data sources characterising 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 from the M25 motorway in London. We use it to train models that predict the probability distribution of incident durations, utilising 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 utilise the concept of Shapley values from 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-minute horizons.