Multiview Clustering via Canonical Correlation Analysis

Clustering algorithms such as k-means perform poorly when the data is highdimensional. A number of efficient clustering algorithms developed in recent years address this problem by projecting the data into a lower-dimensional subspace, e.g. via principal components analysis (PCA) or random projections, before clustering. Such techniques typically require stringent requirements on the separation between the cluster means. Here we present ongoing work on projection-based clustering that addresses this using multiple views of the data. We use canonical correlation analysis (CCA) to project the data in each view to a lower-dimensional subspace. Under the assumption that the correlated dimensions capture the information about the cluster identities, the separation conditions required for the algorithm to be successful are significantly weaker than those of prior results in the literature. We describe experiments on two domains, (a) speech audio and images of the speakers’ faces, and (b) text and links in Wikipedia articles. We discuss several issues that arise when clustering in these domains, in particular the existence of multiple possible “cluster variables” and of a hierarchical cluster structure.