A recognition method based on collective decision making using systems of regularities of various types

A new recognition method that implies weighted voting on systems of “syndromes,” i.e., subregions of the attribute space where objects of one class dominate, is given. It is a modified version of statistically weighted syndromes developed previously. To find syndromes, it searches for optimal partitions within several models of different levels of complexity. Syndromes to be included in the final set used in collective decision making are selected by the criterion for the partitioning degree of classes and by the parameter related to the complexity of the partitioning model involved. The weighted voting procedure can be interpreted as the convex correction of sets of predictors. The generalizing potential of such procedures is discussed. Experimental results of comparing the given method with the previous version (SWS) and alternative techniques are presented. To estimate the efficiency, several criteria are used, including a way to analyze recognition accuracy on the totality of all possible decision rules (ROC analysis).