Spectral clustering using the kNN-MST similarity graph

Abstract

Spectral clustering is a technique that uses the spectrum of a similarity graph to cluster data. Part of this procedure involves calculating the similarity between data points and creating a similarity graph from the resulting similarity matrix. This is ordinarily achieved by creating a k-nearest neighbour (kNN) graph. In this paper, we show the benefits of using a different similarity graph, namely the union of the kNN graph and the minimum spanning tree of the negated similarity matrix (kNN-MST). We show that this has some distinct advantages on both synthetic and real datasets. Specifically, the clustering accuracy of kNN-MST is less dependent on the choice of k than kNN is.