Training dataset censoring

The proposed method of compactness increasing is based on the new measure of similarity between objects - function of rival similarity (FRiS-function) - which allows to describe any type of probability distribution with the set of standards. One can estimate contribution of every object of the dataset into compactness of its class, calculate the quantitative measure of compactness of each class separately and compactness of the whole dataset. As well objects, which influence negatively on the compactness value, can be selected. Main idea of proposed method of training dataset censoring consists in removing such objects. As a result the decision rule, constructed on censored dataset, has a better recognition quality. The set of excluded objects is detected automatically. Effectiveness of the censoring algorithm is illustrated by a model task of two classes recognition.

Keywords

Authors

References