Discriminative Topic Sparse Representation for Text Categorization

In text categorization, feature representation for dimensionality reduction is a key step. Usually, some commonly used methods, e.g., latent semantic analysis (LSA), yield a dense representation or a dense transformation matrix, which is difficult to precisely characterize the document-topic or the topic-word relationship. This paper proposes a novel discriminative topic sparse representation (DTSR) approach for text categorization, in which two stages are included: the topic dictionary construction and sparse representation. Firstly, a discriminative and interpretable dictionary is constructed to characterize the topic-word relationship. The dictionary contains all category center vectors as well as some semantic topic vectors generated by a latent Dirichlet allocation (LDA) model. Furthermore, each document can be represented with a sparse form to obtain a good document-topic relationship. Experimental results on well-known benchmark datasets indicate that the proposed method not only achieves a satisfactory classification performance but also provides a reasonable sparse semantic meaningful.

I. Introduction

With the rapid growth of online digital text data, text categorization has become one of the key techniques for handling and organizing text data [1], [2], in which a major difficulty is the high dimensionality of the input feature space. Feature representation for dimensionality reduction is an effective approach to tackle this problem [3], [4].