ABSTRACT

This chapter discusses approaches for Semisupervised clustering (SSC) based on pointwise and pairwise semisupervision for feature-based representations and related clustering algorithms. It describes semisupervised graph-cuts and spectral clustering. The chapter introduces the GLP formulation, presents a unified view of existing label propagation methods, and illustrates their relationship to semisupervised graph-based clustering. It discusses semisupervised manifold embedding and a set of embedding-based label propagation methods. SSC has become an important part of data mining. The Local and Global Consistency approach gives an alternative graph-based regularization framework for semisupervised learning. The literature on semisupervised clustering can be broadly divided into two families, based on whether one considers feature-based or graph-based representation of the input data. The main idea in cluster kernels is to embed the data into a lower dimensional space based on its cluster structure and then subsequently build a semisupervised learner on the low-dimensional data.