Abstract: In data streams concept is often not stable but change with time. We proposed a selective integration algorithm OSEN(Orientation based Selected ENsemble) for handling concept drift data streams. This algorithm selects a near optimal subset of base classifiers based on the output of each base classifier on validation dataset. Our experiments with synthetic data sets simulating abrupt (SEA) and gradual (Hyperplane) concept drifts demonstrate that selective integration of classifiers built over small time intervals or fixed-sized data blocks can be significantly better than majority voting and weighted voting, which are currently the most commonly used integration technictues for handling concept drift with ensembles. This paper also explained the working mechanism of OSEN from error-ambiguity decomposition. Based on experiments, OSEN improves the generalination ability through reducing the average generalization error of the base classifiers constituting the ensembles.

Key words: Concept drift, Selective ensemble, Naive baycs, Error-ambiguity decomposition