Optimal fusion rules in team classification under three decision structures

In this paper, we study the performance of a team of dichotomous classifiers, where the classifiers' decisions are combined by logical fusion rules. Three decision structures are derived using the confusion matrix of a single classifier and a priori information, and the performances of the different decision structures are compared. First, we consider the performance of a team of three classifiers with a total of 256 fusion rules. Then, we propose a decision structure that utilizes a moderator, i.e., an entity that exploits Bayesian inference from individual classifiers' decisions and makes final decisions based on maximum likelihood classification. We show the benefits of using a moderator (compared to a decision structure without a moderator). Finally, we propose a decision structure that exploits pairing, i.e., fusing the classifiers' decisions sequentially two-by-two. Two pairing schemes, i.e., incremental and tournament-like, are proposed and we show that incremental pairing is the most effective decision structure among the proposed ones.