An l1-version of the spectral clustering algorithm to promote sparse eigenvectors basis
vendredi 13 décembre 2019, 9h30 - 10h30
We propose a variant of the well known spectral clustering algorithm to
address the challenge of finding k underlying communities of a
deterministic graph with an exact community structure whose edges have
been perturbed by a random graph model.
This new procedure, called l1-spectral clustering, is based on the
introduction of a penalty term to find a sparse eigenvectors basis
corresponding to the indicator vectors of each community.
The advantages of the algorithm, compared to the traditional one, are
demonstrated using simulations on synthetic and real data.