Sensitivity Analysis of Machine Learning-Based Models for the Prediction of Temporal Distribution of Traffic Demand: A Case Study from Adelaide, Australia

Abstract

The application of machine learning (ML) models has gained substantial popularity in recent years, extending across diverse industries and areas of practice. Within transport and traffic modelling and analysis, the strong predictive power of ML-based models presents a compelling alternative to traditional prediction and forecasting approaches. This paper explores insights obtained from interpreting ML-based predictive models for the temporal distribution of origin-destination (OD) traffic demand, utilising the methods of Feature Importance and Partial Dependence Plot (PDP) with data from the Adelaide metropolitan area. Findings reveal that while ML-based predictions may not possess the intuitive clarity of conventional approaches (e.g., theory-driven choice modelling), applying robust interpretation methods yields valuable insights that traditional techniques may fail to uncover.