Creating ensemble classifiers through order and incremental data selection in a stream

This paper presents an original time-sensitive traffic management application for road safety diagnosis in signalized intersections. Such applications require to deal with data streams that may be subject to concept drift over various time scales. The method for road safety analysis relies on the estimation of severity indicators for vehicle interactions based on complex and noisy spatial occupancy information. An expert provides imprecise labels based on video recordings of the traffic scenes. In order to improve the performance—overall and for each class—and the stability of learning in a stream, this paper presents new ensemble methods based on incremental algorithms that rely on their sensitivity to the processing order of instances. Different data selection criteria, many used in active learning methods, are studied in a comprehensive experimental evaluation, including benchmark datasets from the UCI machine learning repository and the prediction of severity indicators. The best performance is obtained with a criterion that selects instances which are misclassified by the current hypothesis. The proposed ensemble methods using this criterion and AdaBoost have similar principles and performance, while the proposed methods have a smaller computational training cost.