Drivers' route choice model is essential in transportation software such as navigation, fleet management, and simulation, where the random utility models (RUM) have dominated for years. The authors investigate here whether machine learning (ML) models could be applied into this field, and whether these approaches outperform the traditional models in goodness-of-fit and prediction. The application framework and data structure are proposed, where the challenging problems lie in: (i) to pool data from multiple origin-destination pairs; and (ii) to interpret results for behaviour analysis. All RUM and ML models are applied in a real network. Results suggest that the random forest, one of the ML models, has satisfying performances with acceptable computation time, making it suitable for large network and real-time analysis. This study shows that the ML models can be adopted for behaviour analysis, such as to prioritise the importance of variables, compute the elasticity, and forecast for scenarios. Future directions on combining the RUM and ML models are discussed.