Fast adaptive stacking of ensembles

This paper presents a new ensemble method for learning from non-stationary data streams. In these situations, massive data are constantly generated at high speed and their target function can change over time. The proposed method, named Fast Adaptive Stacking of Ensembles (FASE), uses a meta-classifier to combine the predictions from the base classifiers in the ensemble. FASE maintains a set of adaptive learners, in order to deal with concept drifting data. The new algorithm is able to process the input data in constant time and space computational complexity. It only receives as parameters the confidence level for the change detection mechanism and the number of base classifiers. These characteristics make FASE very suitable for learning from non-stationary data streams. We empirically compare the new algorithm with various state-of-the-art ensemble methods for learning in non-stationary data streams. We use a Naive Bayes classifier and a Perceptron to evaluate the performance of the algorithms over real-world datasets. The experiment results show that FASE presents higher predictive accuracy in the investigated tasks, being also able to bound its computational cost.