In mobility research, demand can be analysed through either theory-based or data-driven methods. Discrete choice models (DCM), offer the reliability of well-defined behavioural assumptions and utility functions. In contrast, machine learning models are more flexible since they require no specific forms for the utility functions and can handle a wider range of data. This study compares the results obtained with DCMs and Neural Networks (NNs), by analysing modal choices in terms of elasticities and Value of Travel Time (VTT). Different NN architectures and two datasets are employed, differing in the number of observations, territorial context, survey year, and data collection methodology. The main contribution of this research is a method for estimating the VTT when employing NNs, also applicable to econometric models. The proposed approach approximates derivatives through a finite difference ratio, a technique previously rarely applied to VTT estimation with NNs. The results show that NNs, in addition to their predictive capability, can reliably recover economically meaningful indicators as well.
Using artificial neural networks to estimate the value of travel time in the context of mode choice modelling
Tuveri G.
;Piras F.;Sottile E.;Meloni I.
2026-01-01
Abstract
In mobility research, demand can be analysed through either theory-based or data-driven methods. Discrete choice models (DCM), offer the reliability of well-defined behavioural assumptions and utility functions. In contrast, machine learning models are more flexible since they require no specific forms for the utility functions and can handle a wider range of data. This study compares the results obtained with DCMs and Neural Networks (NNs), by analysing modal choices in terms of elasticities and Value of Travel Time (VTT). Different NN architectures and two datasets are employed, differing in the number of observations, territorial context, survey year, and data collection methodology. The main contribution of this research is a method for estimating the VTT when employing NNs, also applicable to econometric models. The proposed approach approximates derivatives through a finite difference ratio, a technique previously rarely applied to VTT estimation with NNs. The results show that NNs, in addition to their predictive capability, can reliably recover economically meaningful indicators as well.
I metadati presenti in IRIS UNICA sono rilasciati con licenza Creative Commons CC0 1.0 Universal, mentre i file delle pubblicazioni sono protetti da diritto d'autore, salvo diversa indicazione.
