MMSVC: An efficient unsupervised learning approach for large-scale datasets

We propose a multi-scale, hierarchical framework to extend the scalability of support vector clustering (SVC). Based on multi-sphere support vector clustering, the clustering algorithm called multi-scale multi-sphere support vector clustering (MMSVC) works in a coarse-to-fine and top-to-down manner. Given one parent cluster, the next learning scale is generated by a secant-like numerical algorithm. A local quantity called spherical support vector density (sSVD) is proposed as a cluster validity measure to describe the compactness of the cluster. It is used as a terminate term in our framework. When dealing with large-scale dataset, our method benefits from the easy parameters tuning (robustness of parameters with respect to the clustering result) and the learning efficiency. We took 1.5 million tiny images to evaluate the method. Experimental result demonstrated that our method greatly improved the scalability and learning efficiency of support vector clustering.