Pairwise combination of classifiers for ensemble learning on data streams

This work presents two different voting strategies for ensemble learning on data streams based on pairwise combination of component classifiers. Despite efforts to build a diverse ensemble, there is always some degree of overlap between component classifiers models. Our voting strategies are aimed at using these overlaps to support ensemble prediction. We hypothesize that by combining pairs of classifiers it is possible to alleviate incorrect individual predictions that would otherwise negatively impact the overall ensemble decision. The first strategy, Pairwise Accuracy (PA), combines the shared accuracy estimation of all possible pairs in the ensemble, while the second strategy, Pairwise Patterns (PP), record patterns of pairwise decisions during training and use these patterns during prediction. We present empirical results comparing ensemble classifiers with their original voting methods and our proposed methods in both real and synthetic datasets, with and without concept drifts. Our analysis indicates that pairwise voting is able to enhance overall performance for PP, especially on real datasets, and that PA is useful whenever there are noticeable differences in accuracy estimates among ensemble members, which is common during concept drifts.