TY - GEN UR - https://www.events.tum.de/frontend/index.php?sub=94 A1 - Hillel, Tim A1 - Bierlaire, Michel A1 - Elshafie, Mohammed A1 - Jin, Ying EP - 5 AV - public N2 - Predicting passenger mode choice is an essential task for transport network simulation and operations management. Machine Learning (ML) classifiers trained on trip diary data are increasingly being investigated as an alternative to Discrete Choice Models (DCMs) for mode choice prediction. In order to determine the suitability of a particular algorithm for predicting mode choice, it is crucial to be able to reliably quantify its predictive performance for unknown trips. In this paper we propose and test a new framework for performance estimation of classification algorithms. Firstly, we address three key issues in the literature relating to a) performance metrics for model training, b) sampling methods for model validation, and c) hyperparameter selection for model optimisation. Alternative methods are proposed for each of these issues to form the new framework. The framework is then used to compare the suitability of eight different ML classification algorithms for predicting mode choice. Finally, the implications of the sampling method employed are then investigated experimentally. ID - discovery10173006 CY - Athens, Greece N1 - This version is the version of record. For information on re-use, please refer to the publisher?s terms and conditions. PB - hEART SP - 1 TI - A new framework for assessing classification algorithms for mode choice prediction Y1 - 2018/09/07/ T3 - Symposium of the European Association for Research in Transportation ER -