Ensemble based data stream mining with recalling and forgetting mechanisms

Using ensemble of classifiers on sequential chunks of training instances is a popular strategy for data stream mining. Aiming at the limitations of the existing approaches, we introduce recalling and forgetting mechanisms into ensemble based data stream mining, and put forward a new algorithm MAE (Memorizing based Adaptive Ensemble) to mine chunk-based data streams with concept drifts. Ensemble pruning is used as a recalling mechanism to select useful component classifiers for each incoming data chunk. Ebbinghaus forgetting curve is adopted as a forgetting mechanism to evaluate and replace the component classifiers in the memory repository. Experiments have been performed on datasets with different types of concept drifts. Compared with traditional ensemble approaches, the results show that MAE is a good algorithm with high and stable accuracy, less predicting time and moderate training time.