Co-cluster Structure Visualization by Spectral Ordering and Its Characteristics

Abstract

Cluster structure analysis is often utilized as the basic step in unsupervised classification. In this paper, a spectral ordering-based visual cluster assessment method for relational data analysis is modified so that it can be applied to co-cluster structure analysis in cooccurrence information data. Object-by-item cooccurrence information is first enlarged into an (object + item)-by-(object + item) relational data matrix, and then, co-cluster structure is visually assessed through simultaneous ordering of objects and items in the enlarged matrix. Additionally, due to the sparse nature of the enlarged matrix, the computational cost can be decreased with the eigen problem of a reduced matrix. The characteristic features of the proposed approach are demonstrated through several numerical experiments including social analysis of Japanese prefectural statistics.