Non-negative Sparse Semantic Coding for text categorization

In text categorization, the dimensionality reduction methods, such as latent semantic indexing and nonnegative matrix factorization, commonly yield the dense representation that is not consistent with our common knowledge. On the other hand, the popular sparse coding methods are time-consuming and their dictionaries might contain negative entries, which is difficulty to interpret the semantic meaning of text. This paper proposes a novel Non-negative Sparse Semantic Coding (NSSC) approach for text reprentation. NSSC provides an efficient algorithm to construct a set of non-negative basis vectors that span a low dimensional semantic subspace, where each document obtains a non-negative sparse representation corresponding to these basis vectors. Extensive experimental results have shown that the proposed approach achieves a good performance and presents more interpretability with respect to these semantic concepts.