New diversity measure for data stream classification ensembles.

The diversity of a voting committee is one of the key characteristics of ensemble systems. It determines the benefits that can be obtained through classifier fusion. There are many measures of diversity that can be used in classical decision-making systems which operate in stationary environments. A plethora of algorithms have also been proposed to ensure ensemble diversity. Bagging and boosting are a few of the most popular examples. Unfortunately, these measures and algorithms cannot be applied in systems that process streaming data. Not only must a different implementation be designed for processing fast moving samples in a stream, but the notion of diversity must also be redefined. In this paper it is proposed to assess diversity based on analysis of classifier reactions to changes in data streams. Therefore, two novel error trend diversity measures are introduced that compare the error trends of classifiers while processing subsequent samples. A practical application of these measures is also proposed in the form of a novel error trend diversity driven ensemble algorithm, where our measures are incorporated into the training procedure. The performance of the proposed algorithm is evaluated through a series of experiments and compared to several competing methods. The results demonstrate that our measures accurately evaluate diversity and that their application facilitates the creation of small and effective ensemble classifier systems.