The main idea of spectral clustering is in modeling of a data set in a form of a simple undirected weighted graph, and observing the graph Laplacian spectrum. Usually, data partition can be reconstructed from dominant eigenvectors. We introduce several heuristic algorithms for accurate determination of number of clusters. The algorithms are based on properties of coupling matrix introduced in [1]. In this presentation we give several examples of datasets sucessfully clustered by our algorithms, both artificial and real-world ones.

Matematika