Grouping Training Samples To Increase The Validity Of Recognition Algorithm Solutions

Abstract

The paper is devoted to increasing the validity of recognition algorithms by dividing training samples into subsets of their related instances and using such subsets at the stage of training recognition procedures and in the criteria for making decisions about the types of signals considered. Software tools for implementing this approach are proposed. A variant of grouping training samples is considered, which is based on dividing the samples into clusters of similar instances. A supposed possible increase in the validity of the algorithm solutions here is associated with a more detailed consideration of the dislocations of the accumulation of signals in feature spaces. The proposed system for forming a system of isolines and shape characteristics can be used in algorithms for recognizing pathological patterns in the ECG signal shape. This approach also implements the principle of data dimensionality reduction about the characteristics of the waveform, which can be used in deep learning systems. The introduction reveals the general formulation and relevance of the proposed research. Section 1 discusses general issues of problem formulation: specific types of recognized signals, features for their description, the recognition algorithm chosen for research, the procedure for assessing the validity of its solutions and assessing the initial validity of solutions for the original algorithm before grouping training samples. In Section 2, we study the grouping of training samples of the algorithm by their separate clustering according to the types of recognized signals. An assessment of its effectiveness is given. The results of the work are briefly disclosed in the general conclusions.