Machine learning algorithms for analysis and categorization of electronic text documents

Abstract

The paper considers two main algorithmic problems arising in the field of analysis and categorization of electronic text documents. The first problem is the development of efficient formal representation models of hypertext data. Such models should incorporate both the text content of the documents and their hyperlink structure. The second problem is the development of efficient multi-label classification algorithms. The multi-label classification is a classification in the case of essentially overlapping classes when any classifying object (document) may belong to several classes (topics) simultaneously. A new text document representation model based on the extraction of frequent episodes of lexemes (or N-grams) and a new method for the hyperlinks structure analysis based on N-gram classification of a URL-hyperlink text are proposed. In this algorithm the lexemes are replaced in the initial document text by special tags. A possibility of applying the «one-versus-one» decomposition approach for solving the multi-label classification problem is also studied. A new multi-label classification method based on the paired comparison approach is proposed. In this method each pair of possibly overlapping classes is separated by two probabilistic binary classifiers that isolate the overlapping and non-overlapping areas. Then, the individual probabilities generated by the binary classifiers are combined together to estimate the final class probabilities that fit to the extended Bradley-Terry model with ties. In the proposed method, nonrelevant classes are cut off by a specially developed threshold function defined in the class-relevance vector space. Experimental performance evaluation using two well-known multi-label benchmark datasets demonstrated the higher-accuracy results of the proposed method as compared to existing approaches.