Semantic concept space based progressive transductive learning for text classification

Though the learning ability of Progressive Tranductive Support Vector Machines may be independent of the dimensionality of feature space, dimension reduction is an essential issue to efficiently handle large scale terms in practical applications of text classification. A novel approach is proposed to reduce dimensions of documents and extract semantic concepts of words using Kernel Principle Component Analysis. The better performance are achieved via indirectly compute and store covariance matrix at each iteration. The experiment results of Reuters -21578 data show substantial improvements over progressive inductive methods in semantic concept space. It is very promising for a mixed training set of small labeled examples and a large number of unlabeled examples.