Joint of local and global structure for clustering

Abstract

We consider the general problem of clustering from labeled and unlabeled data, which is often called semi-supervised learning or transductive inference. Most of existing works consider the intrinsic local or global structure of the dataset, which introduced poor clustering performance in real case scenarios. In this paper, we study the complementary relationship between local and global structure of a dataset, and proposed to obtain a better clustering performance via label propagation process. To validate our proposed method, we conduct experiment on the two-moon problem, and find that our method yields encouraging experimental results on a number of classification problems and demonstrates effective use of unlabeled data.