RankBased Multiple Classifier Decision Combination A Theoretical Study.pdfVIP

Rank-Based Multiple Classifier Decision Combination: A Theoretical Study Afs¸ar Saranlı and M¨ubeccel Demirekler Speech Processing Laboratory Department of Electrical and Electronics Engineering Middle East Technical University, ˙ ˙ Ismet In¨on¨u Bulv. 06531, Ankara, T¨urkiye. fafsars, demirekg@metu.edu.tr Abstract show that the methods of combining multiple classifiers is of considerable interest in many diverse applications of This study presents a theoretical investigation of the rank- pattern recognition. based multiple classifier decision problem for closed-set Xu and his colleagues have categorized multiple classifier pattern identification. The problem of combining the de- combination systems with respect to the type of raw out- cisions of more than one classifiers with raw outputs in put information from each classifier, resulting in three such the form of candidate class rankings is considered and for- categories [29]. These are categories where the classifier mulated as a general discrete optimization problem with outputs are single class labels (Type 1), rankings of a sub- an objective function based on the total probability of cor- set of candidate classes from highest to lowest “likelihood” rect decision. This formulation uses certain performance (Type 2) and complete set of measurement values for the statistics about the joint behavior of the ensemble of classi- candidate clas