Recognizing bounds of context change in on-line learning

The on-line algorithms in machine learning are intended to discover unknown function of the domain based on incremental observing of it instance by instance. These algorithms have a great ability for adaptation to each new situation they appear in time, or to a new context. In this paper we propose an algorithm for identifying the moments when the context of the 'environment' changes. The key idea is that if the predictor, which was based on some previous examples, becomes crucially inconsistent with new examples then we identify context change at the placement of this instance. If most of the predictors from diverse ensemble start to identify context change then we report total context change. We experimentally illustrate that this idea works very well on Vowel recognition dataset.