Semi-Supervised Spectral Clustering

Xiaoyi Mai, Romain Couillet

Abstract: In this article, we propose a semi-supervised version of spectral clustering, a widespread graph-based unsupervised learning method. The semi-supervised spectral clustering has the advantage of producing consistent classification of data with sufficiently large number of labelled or unlabelled data, un-like classical graph-based semi-supervised methods which are only consistent on labelled data. Theoretical arguments are provided to support the proposition of this novel approach, as well as empirical evidence to confirm the theoretical claims and demonstrate its superiority over other graph-based semi-supervised methods.