An incremental subspace learning algorithm to categorize large scale text data

Abstract

The dramatic growth in the number and size of on-line information sources has fueled increasing research interest in the incremental subspace learning problem. In this paper, we propose an incremental supervised subspace learning algorithm, called Incremental Inter-class Scatter (US) algorithm. Unlike traditional batch learners, US learns from a stream of training data, not a set. US overcomes the inherent problem of some other incremental operations such as Incremental Principal Component Analysis (PCA) and Incremental Linear Discriminant Analysis (LDA). The experimental results on the synthetic datasets show that US performs as well as LDA and is more robust against noise. In addition, the experiments on the Reuters Corpus Volume l (RCV1) dataset show that US outperforms state-of-the-art Incremental Principal Component Analysis (IPCA) algorithm, a related algorithm, and Information Gain in efficiency and effectiveness respectively. © Springer-Verlag Berlin Heidelberg 2005.