Appraising Machine and Deep Learning Techniques for Traffic Conflict Prediction with Class Imbalance

Abstract

Predicting traffic conflicts is pivotal for vehicle-based active safety system to prevent crashes. Yet, conflict prediction is a challenging task as correct prediction depends on the nature of data and techniques employed. Moreover, traffic conflicts data are naturally imbalanced with traffic conflicts being the minority class. Working with imbalanced dataset might result in biased and inaccurate predictions. Therefore, this study aims to appraise machine learning and deep learning techniques systematically, to identify the optimal technique which can reliably predict real-time traffic conflicts by making use of cost-sensitive learning. Five machine learning techniques were optimised and utilised including: logistic regression (LR), Support vector machines (SVM), deep neural networks (DNN), long short-term memory (LSTM) and LSTM convolutional neural network (LSTM-CNN) to appraise their predictability performance using a large, imbalanced, and disaggregated traffic dataset. Unlike existing studies, a wide range of interconnected factors are employed for real-time traffic conflict prediction to provide a more reliable prediction outcome. A large heterodox dataset was gathered from the M1 motorway in the UK to evaluate these techniques. Results suggested that DNN outperform other techniques in predicting conflicts with 0.72 sensitivity at 5% false alarm rate. Such promising results reflect that DNNs can be further applied to deepen our understanding in predicting traffic conflicts design more reliable primary safety systems for intelligent vehicles. Moreover, exploring state-of-the-art classification techniques with class imbalance on big data is significant to the future of big data analytics.