Class Selection Based Iterative Supervised Latent Semantic Indexing for Text Categorization

Latent semantic indexing (LSI) is an effective technique for feature extraction in text mining, and supervised LSI (SLSI) algorithms have been proposed to exploit the class labels of training data. In this paper, we propose an iterative SLSI framework based on class selection. We show that a previous iterative SLSI algorithm is an instance of the framework. We also propose a method under our framework, which selects a class at each iteration using a simple classifier and computes the main bias vector of one class only. Our experiments demonstrate that the proposed method both improves the classification accuracy and reduces the computation cost.