A mixture model approach to spectral clustering and application to textual data

The spectral clustering algorithm is a technique based on the properties of the pairwise similarity matrix coming from a suitable kernel function. It is a useful approach for high-dimensional data since the units are clustered in feature space with a reduced number of dimensions. In this paper, we consider a two-step model-based approach within the spectral clustering framework. Based on simulated data, first, we discuss criteria for selecting the number of clusters and analyzing the robustness of the model-based approach concerning the choice of the proximity parameters of the kernel functions. Finally, we consider applications of the spectral methods to cluster five real textual datasets and, in this framework, a new kernel function is also proposed. The approach is illustrated on the ground of a large numerical study based on both simulated and real datasets.