Semi-supervised Multiclass Clustering Based on Signed Total Variation
Peter Berger, Thomas Dittrich, Gabor Hannak, Gerald Matz
Abstract
We consider the problem of semi-supervised clustering for multiple (more than two) classes. The proposed clustering algorithm uses the (dis)similarity of given data to learn the unknown cluster labels. We quantify label (dis)similarity in terms of the new concept of signed total variation (TV). The clustering task is formulated as a convex optimization problem with an ℓ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> -norm regularization term that helps when only few labels are known. We solve the optimization problem by developing an ADMM-based algorithm whose per-iteration complexity scales linearly with the number of edges and the number of clusters. Our algorithm admits a distributed implementation and can therefore efficiently handle large-dimensional problems. Numerical experiments demonstrate the superiority of our scheme.
BibTeX
@inproceedings{icassp2019_semisupervisedmu,
title = {Semi-supervised Multiclass Clustering Based on Signed Total Variation},
author = {Peter Berger and Thomas Dittrich and Gabor Hannak and Gerald Matz},
booktitle = {ICASSP 2019},
year = {2019}
}