BUILDING MULTI-CLASSIFFIER SYSTEMS WITH ANT COLONY OPTIMIZATION

Abstract

n recent years, the development of multi-classifier systems has become an active research field. A multi-classifier system is an
ensemble of classification algorithms whose individual outputs are fused together for better accuracy and interpretability. An
important aspect when designing such systems is related to the heterogeneity of the building blocks (classifiers) that make up the
ensemble, since previous studies have uncovered that a more diverse ensemble often boosts up the overall classification power.
Some statistical measures can be used to estimate how diverse the classifier ensembles are; they are called diversity measures.
Another issue to be considered is the number of individual classifiers included in the model: the lower the number of classifiers,
the simpler the resulting system. In general terms, the parsimony principle is highly desired in such ensembles, since a bulky
ensemble will also be a very time-consuming model. Finding the minimal subset of individual classifiers that brings about the
best system performance can be posed as a combinatorial optimization problem. In this paper, we address the problem of
building multi-classifiers systems from the perspective of Ant Colony Optimization (ACO), a widely popular and effective
metaheuristic optimization algorithm. The main reason behind the use of ACO lies on its strong ability to solve entangled
combinatorial optimization problems. An empirical analysis is included to statistically validate the benefits of our proposal.