Dimensionality Reduction and Visualization by Doubly Kernelized Unit Ball Embedding

In this paper, we present a nonlinear dimensionality reduction algorithm which is aimed to preserve the local structure of data by building and exploiting a neighborhood graph. The cost function is defined to minimize the discrepancy between the similarities of points in the input and output spaces. We propose an effective way to calculate the input and output similarities based on Gaussian and polynomial kernel functions. By maximizing the within-cluster cohesion and between-cluster separation, KUBE remarkably improves the quality of clustering algorithms on the low-dimensional embedding. Our experiments on image recognition datasets show that KUBE can learn the structure of manifolds and it significantly improves the clustering quality.