Parameter identification in Choquet Integral by the Kullback-Leibler divergence on continuous densities with application to classification fusion

Abstract

Classifier fusion is a means to increase accuracy and decision-making of classification systems by designing a set of basis classifiers and then combining their outputs. The combination is made up by non linear functional dependent on fuzzy measures called Choquet integral. It constitues a vast family of aggregation operators including minimum, maximum or weighted sum. The main issue before applying the Choquet integral is to identify the 2 M 2 parameters for M classifiers. We follow a previous work by Kojadinovic and one of the authors where the identification is performed using an informationtheoritic approach. The underlying probability densities are made smooth by fitting continuous parametric and then the Kullback-Leibler divergence is used to identify fuzzy measures. The proposed framework is applied on widely used datasets.