Misuno I. Vector representation of information for retrieval and classification tasks

Dissertation is devoted to developing and increasing efficiency of methods for vector representations and processing of visual and textual information in search and classification tasks. Methods for selection of binary features (elements of vector data representation) based on criteria of information and redundancy criteria have been developed. In classification tasks they provide computational savings preserving classification quality. Methods for increasing classification quality of data with vector representation have been developed, including elaboration of the perceptron rule, extending training set, combining multiple classification results and selection of reliable results, as well as rejecting unreliable classification results. Methods for distributed representation of text information taking into account semantic similarity have been developed that exploit context vectors with discrete elements formed using documents from a training collection as contexts. Efficiency of developed methods was experimentally confirmed on real-world data. Software Neurocomputer SNC has been developed that provide efficient instrumental tools for development, investigation, and practical application of methods for intelligent data processing.