Analysis of the Correlation Between Majority Voting Error and the Diversity Measures in Multiple Classifier Systems

Abstract

Combining classifiers by majority voting (MV) has
\nrecently emerged as an effective way of improving
\nperformance of individual classifiers. However, the
\nusefulness of applying MV is not always observed and
\nis subject to distribution of classification outputs in a
\nmultiple classifier system (MCS). Evaluation of MV
\nerrors (MVE) for all combinations of classifiers in MCS
\nis a complex process of exponential complexity.
\nReduction of this complexity can be achieved provided
\nthe explicit relationship between MVE and any other
\nless complex function operating on classifier outputs is
\nfound. Diversity measures operating on binary
\nclassification outputs (correct/incorrect) are studied in
\nthis paper as potential candidates for such functions.
\nTheir correlation with MVE, interpreted as the quality
\nof a measure, is thoroughly investigated using artificial
\nand real-world datasets. Moreover, we propose new
\ndiversity measure efficiently exploiting information
\ncoming from the whole MCS, rather than its part, for
\nwhich it is applied.