Self-Tuning Semi-Supervised Spectral Clustering

Abstract

Spectral clustering (SC), as an unsupervised learning algorithm, has been used successfully in the field of computer vision for data clustering. In some applications, however, background prior knowledge can be easily obtained, such as pairwise constraints. Therefore, semi-supervised learning is getting increasing attention in recent years. In this paper, a new algorithm named self-tuning semi-supervised spectral clustering (STS <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> C) is proposed. We incorporate two types of instance-level constraints-must-link and cannot-link into SC and use self-tuning parameter to solve the scaling parameter selection problem in SC. Experimental results over four datasets from UCI machine learning repository show that STS <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> C performs better than semi-supervised spectral clustering with fixed scaling parameter, and also avoids the time-consuming procedure of parameter selection.