Document Clustering: Before and After the Singular Value Decomposition

Document Clustering is an issue of measuring similarity between documents and grouping similar documents together. Information Retrieval (IR) is an issue of comparing query with a collection of documents to locate a set of documents relevant to a particular query. In the vector space IR model, a query is treated as a document which consists of a few terms. Therefore, in both clustering and retrieval we necessarily address issues involving representation of documents and computation of similarities between a set of documents. In the vector space IR model, term-document matrix is computed from a collection of documents using a certain weighting scheme. Latent Semantic Indexing, an efficient vector space retrieval approach, uses Singular Value Decomposition technique to reduce the rank of the original term-document matrix. Theoretically, SVD, a dimensionality reduction technique, performs a term-to-concept mapping, and therefore, conceptual indexing and retrieval is made possible. In this paper, we discuss the clustering of documents by calculating pair-wise similarity between documents using the original term-document matrix and the decomposed term-document matrix. We also report an evaluation method based on the clustering hypothesis and analyze the clustering results.