Prediction of destination choice in transport modelling with mobile phone data and machine learning models: Experiences from Tønsberg
Abstract
The development of transportation infrastructure is a prolonged and complex process that requires in depth planning. The planning process is costly, and further investments for construction of infrastructure is even larger. Transport modelling predicts future traffic flow to appropriately dimension the required infrastructure. The study addresses the prediction of destination choice in the four-step model (4SM) in transport modelling with combining mobile phone data and machine learning algorithms.
Traditionally, destination choice models relied on theory-based discrete choice models using data from travel surveys. However, emerging of Big Data and advances in artificial intelligence have facilitated the use of more representative data, possible for enhancing predictive accuracy and potentially reducing costs and risks associated with the over- or under-construction of infrastructure. Previous research has focused on either the application of mobile phone data or artificial intelligence independently in transport modelling. This research aims to investigate the performance by combining these two novel approaches in transport modelling.
The selection of algorithms tested includes Logistic Regression, Support Vector Machine, Random Forest and Naïve Bayes. They are tested within a basic analytical pipeline to assess performance. Performance evaluations were conducted using a five-fold cross-validation on performance metrics, and through a multi-criteria decision analysis (MCDA) to consider both qualitative and quantitative criteria for model performance. Results indicate robust performance across all algorithms on mobile phone data, with Random Forest performing best, considering both quantitative and qualitative metrics, achieving a prediction accuracy of 0.943. Naïve Bayes and Support Vector Machine followed with 0.925 and 0.895, respectively, while Logistic Regression achieved 0.832. Simpler hyperparameters yielded better results for Support Vector Machine and Logistic Regression, whereas Random Forest utilized more complex hyperparameters for best performance.
These findings suggest that integrating mobile phone data with machine learning algorithms holds substantial promise for enhancing the prediction of destination choices. Future research should explore the model’s applicability across different geographic contexts and further steps towards its implementation and deployment.
Keywords: transport modelling, travel demand, mobile phone data, machine learning, artificial intelligence, Big Data, destination choice.