A Multi-agent Ensemble of Classifiers

Abstract

It is well-known that every classifier method or algorithm, being Multi-Layer Perceptrons, Decisions Trees or the like, are heavily dependent on data. That is to say, their performance varies significantly whether training data is balanced or not, multi-class or binary, or if classes are defined by numeric or symbolic variables. Some unwanted issues arise, for example, classifiers might be over-trained, or they could present bias or variance, all of which lead to poor performance. The classifiers performance can be analyzed by metrics such as specificity, sensitivity, F-Measure, or the area under the ROC curve. Ensembles of Classifiers are proposed as a means to improve classifications tasks. Classical approaches include Boosting, Bagging and Stacking. However, they do not present cooperation among the base classifiers to achieve a superior global performance. For example, it is desirable that individual classifiers are able to communicate each other what tuples are classified correctly and which are not so errors are not duplicated. We propose an Ensemble of Classifiers that relies on a cooperation mechanism to iteratively improve the performance of both, base classifiers and ensemble. Information Fusion is used to reach a decision. The ensemble is implemented as a Multi-Agent System (MAS), programmed on the JADE platform. The base classifiers are taken from WEKA, as well as the calculation of the performance metrics. We prove the ensemble with a real dataset that is unbalanced, multi-class, and high-dimensional, obtained from a psychoacoustics study.