Predicting choices of street-view images: A comparison between discrete choice models and machine learning models

Abstract

Recently, there has been a growing interest in comparing machine learning models and Discrete Choice Models. However, no studies have been conducted on image choice problems. This study aims to fill this gap by conducting a stated preference experiment that involves choosing streets for cycling based on real-world street-view images. The choice data obtained were used to estimate and compare four models: Multinomial Logit, Mixed Logit, Deep Neural Network, and Convolutional Neural Network. Additionally, the study tested the effects of different data formats on the models' performances, including semantic interpretation, semantic segmentation, raw image, semantic map, and enriched image. The comparison focused on the models' explainability and out-of-sample predictability with new but similar choice data. The results show that (1) the Discrete Choice Models exhibit nearly equal predictability to the Deep Neural Network models, but significantly outperform the Convolutional Neural Network models; (2) the Discrete Choice Models are more explainable than the Deep Neural Network models; and (3) models trained on semantic interpretation data demonstrate better predictability than those trained on semantic segmentation data and imagery data.