A Comparison on How Statistical Tests Deal with Concept Drifts

RCD is a framework proposed to deal with recurring concept drifts. It stores classifiers together with a sample of data used to train them. If a concept drift occurs, RCD tests all the stored samples with a sample of actual data, trying to verify if this is a new context or an old one that is recurring. This is performed by a non-parametric multivariate statistical test to make the verification. This paper describes how two statistical tests (KNN and Cramer) can distinguish between new and old contexts. RCD is tested with several base classifiers, in environments with different rates-of-change values, with gradual and abrupt concept drifts. Results show that RCD improves single classifiers accuracy independently of the statistical test used.