Monitoring a Dynamic Weighted Majority Method Based on Datasets with Concept Drift

Abstract

Monitoring changes during a learning process is an interesting area of research in several online applications. The most important problem is how to detect and explain these changes so that the performance of the learning model can be controlled and maintained. Ensemble methods have perfectly coped with concept drift. This paper presents an online classification ensemble method designed for concept drift entitled dynamic weighted majority (DWM) algorithm. It adds and removes experts based on their performance and adjusts learner’s weights taking into account their age in the ensemble as well as their historical correct prediction. The idea behind this paper is to monitor the classification error rates of DWM based on a time adjusting control chart which adjusts the control limits each time an adjustment condition is satisfied. Moreover, this paper handles datasets with concept drift and analyzes the impact of the diversity of base classifiers, noises, permutations and number of batches. Experiments tested with ANOVA and confirmed by Tukey’s test have shown that monitoring classification errors with DWM algorithm has a perfect reaction capacity to different types of concept drift.