Recurrent Adaptive Classifier Ensemble for Handling Recurring Concept Drifts

For most real-world data streams, the concept about which is obtained may shift from time to time, a phenomenon known as drift. applications such nonstationary time-series data, drift often occurs in cyclic fashion, and previously seen concepts will reappear, supports unique kind of recurring concepts. A cyclically drifting exhibits tendency return visited states. Existing machine learning algorithms handle by retraining model if detected, leading loss information was well learned model, recur again next phase. common remedy for retain reuse models, but process time-consuming computationally prohibitive environments appropriately select any optimal ensemble classifier capable accurately adapting To learn streaming fast accurate are needed time-dependent applications. Most existing designed do not take into account presence efficiently with minimum computational overheads, we propose novel evolving method called Recurrent Adaptive Classifier Ensemble (RACE). The algorithm preserves an archive models that diverse always trains both new classifiers. empirical experiments conducted on synthetic stream benchmarks show RACE significantly adapts more than some state-of-the-art classifiers based reuse.