Fusing and filtering arrogant classifiers

Abstract

Given a finite collection of classifiers trained on n-class data, one wishes to fuse the classifiers to form a new classifier with improved performance. Typically, the fusion is performed on the output level using logical ANDs and ORs. Sometimes classifiers are arrogant and will classify a feature vector without any prior experience (data) to justify their decision. The proposed fusion is based on the arrogance of the classifier and the location of the feature vector in respect to training data. Given a feature vector x, if any one of the classifiers is an expert on x then that classifier should dominate the fusion. If the classifiers are confused at x then the fusion rule should be defined in such a way to reflect this confusion. If the classifier is arrogant, then its results should not be considered and, thus, filtered out from the fusion process. We give this fusion rule based upon the metrics of veracity and experience.