Adaptive Time Window Size to Track Concept Drift

This paper proposes an approach to track concept drift in order to improve the classifier performance. This approach uses an adaptive time window size in order to detect a drift according to its dynamics (slow/moderate/fast). The goal is to update the classifier using sufficient number of patterns related to environment changes. Since the classifier may misclassify drifted patterns with its old parameters, an expert is asked to provide the true class label for these patterns. This approach is used to detect at early stage a leak in the steam generator of nuclear power generators Prototype Fast Reactors.