Support Vector Clustering: From Local Constraint to Global Stability

Summary

During recent years, numerous kernel based clustering algorithms have been proposed. Support Vector Machines, as standard tools for classification and clustering, have played an important role in this area. Information retrieval from the data in the absence of target labels, as an appealing aspect of unsupervised procedures, necessitates realization of a clear perception of the topological structure of the patterns. Moreover, the learner will require creating a balance between generalization and denoising abilities and regularization of the complexity as well. The Support Vector Clustering (SVC) approach, though globally informative, lacks mechanisms to take advantage of the inferences made from local statistics. Parameterizing the algorithm and embedding local methods is still an open problem in the SVC algorithm. In this chapter, the unsupervised support vector method for clustering is studied. Following the previous works, we will put forward recent efforts aimed at establishing a reliable framework for automating the clustering procedure and regularizing the complexity of the decision boundaries. The novel method takes advantage of the information obtained from a Mixture of Factor Analyzers (MFA) assuming that lower dimensional non-linear manifolds are locally linearly related and smoothly changing.