University of Washington - Department of Statistics
Spectral clustering is a pairwise clustering technique which uses the eigenvectors and eigenvalues of a normalized similarity matrix to cluster the data. While this is a popular clustering method, a limiting factor in spectral segmentation is that the similarity matrix is not usually known a priori. In this talk we present our method for learning the similarity matrix. We introduce the idea of using a clustering quality term, the gap, regularized be a clustering stability term, the eigengap. We use this methodology to learn the similarity matrix such that the resulting clustering is both "good" and "unique." In this talk we will give an introduction to spectral clustering motivated by random walks. We will present our supervised method in detail, which assumes that a training set with known clustering labels is available for learning the similarity matrix. We will also briefly discuss how we extend our methodology to the unsupervised and semi supervised frameworks.