Adapting dynamic classifier selection for concept drift

Abstract One popular approach employed to tackle classification problems in a static environment consists in using a Dynamic Classifier Selection (DCS)-based method to select a custom classifier/ensemble for each test instance according to its neighborhood in a validation set, where the selection can be considered region-dependent. This idea can be extended to concept drift scenarios, where the distribution or the a posteriori probabilities may change over time. Nevertheless, in these scenarios, the classifier selection becomes not only region but also time-dependent. By adding a time dependency, in this work, we hypothesize that any DCS-based approach can be used to handle concept drift problems. Since some regions may not be affected by a concept drift, we introduce the idea of concept diversity, which shows that a pool containing classifiers trained under different concepts may be beneficial when dealing with concept drift problems through a DCS approach. The impacts of pruning mechanisms are discussed and seven well-known DCS methods are evaluated in the proposed framework, using a robust experimental protocol based on 12 common concept drift problems with different properties, and the PKLot dataset considering an experimental protocol specially designed in this work to test concept drift methods. The experimental results have shown that the DCS approach comes out ahead in terms of stability, i.e., it performs well in most cases requiring almost no parameter tuning.