UNSUPERVISED METRIC LEARNING VIA NONLINEAR FEATURE SPACE TRANSFORMATIONS

Pin Zhang, Bibo Shi, JundongLiu

Abstract: In this paper, we propose a nonlinear unsupervised metric learning framework to boost of the performance of clustering algorithms. Under our framework, nonlinear distance metric learning and manifold embedding are integrated and conducted simultaneously to increase the natural separations among data samples. The metric learning component is implemented through feature space transformations, regulated by a nonlinear deformable model called Coherent Point Drifting (CPD). Driven by CPD, data points can get to a higher level of linear separability, which is subsequently picked up by the manifold embedding component to generate well-separable sample projections for clustering. Experimental results on synthetic and benchmark datasets show the effectiveness of our proposed approach over the state-of-the-art solutions in unsupervised metric learning.

TL;DR: a nonlinear unsupervised metric learning framework to boost the performance of clustering algorithms.

Keywords: Metric Learning, K-means, CPD, Clustering