Visual Component Analysis

Abstract

We propose to integrate information visualization techniques with factor analysis. Specifically, a principal direction derived from a principal component analysis (PCA) of the data is displayed together with the data in a scatterplot matrix. The direction can be adjusted to coincide with visual trends in the data. Projecting the data onto the orthogonal subspace allows determining the next direction. The set of directions identified in this way forms an orthogonal space, which represents most of the variation in the data. We call this process visual component analysis (VCA). Furthermore, it is quite simple to integrate VCA with clustering. The user fits poly-lines to the displayed data, and the poly-lines implicitly define clusters. Per-cluster projection leads to the definition of per-cluster components.