Predicting and explaining lane-changing behaviour using machine learning: A comparative study

Predicting lane-changing behaviour is an integral part of lane-changing decision models and has a significant impact on both traffic flow characteristics and traffic safety. A variety of lane-changing decision models have been developed for this purpose, with most of them focussing only on explaining lane-changing behaviour, while assessing the predictive capability of these models has comparatively received less attention. Meanwhile, machine learning techniques are often preferred for prediction purposes, but their application to predicting lane-changing behaviour is limited. However, the lack of interpretability of machine learning techniques is often criticised and needs a solution. Motivated by these research needs, this study explains and predicts driver’s mandatory and discretionary lane-changing behaviours using a set of suitable machine learning techniques. Input features are objectively selected using the technique of Recursive Feature Elimination, and standard classification metrics are employed to select the best model. By accounting for class imbalance, this study finds that the Extra Trees classifier outperforms other machine learning techniques as well as a conventional utility theory-based model in predicting lane-changing behaviour. Furthermore, by keeping the model hyperparameters unchanged, this classifier shows good transferability in predicting lane-changing behaviour when tested on a completely new dataset. Finally, through explainable artificial intelligence, the output of the Extra Trees classifier is interpreted. The findings of this study advocate the use of machine learning techniques in future studies for explaining and predicting lane-changing behaviour.