Learning to detect incidents from noisily labeled data

摘要

Many deployed traffic incident detection systems use algorithms that require significant manual tuning. We seek machine learning incident detection solutions that reduce the need for manual adjustments by taking advantage of massive databases of traffic sensor measurements. We first examine which traffic flow features are most useful for the incident detection task. Then we show that a supervised learner based on the SVM model outperforms a fixed detection model used by state-of-the-art traffic incident detectors. However, the performance of a supervised learner suffers from temporal noise in the data labels due to imperfections of the incident logging procedure. Correcting these misaligned incident times in the training data achieves further improvements in detection performance. We propose a label realignment model based on a dynamic Bayesian network to re-estimate the correct position (time) of the incident in the data. Training on the automatically realigned data then consistently leads to improved detection performance in the low false positive region.