B0387
Title: Optimal nonparametric classification via radial distance
Authors: Akifumi Okuno - Institute of Statistical Mathematics (Japan) [presenting]
Ruixing Cao - Kyoto University (Japan)
Kei Nakagawa - Nomura Asset Manadgement (Japan)
Hidetoshi Shimodaira - Kyoto University and RIKEN AIP (Japan)
Abstract: Conventional kernel is smoother and k-nearest neighbour approaches estimate the label of a query by considering a radial distance (i.e., a distance from the query). While the radial-distance-based approach is applicable to various types of complex data as long as their distance (or pseudo-distance) can be measured, they are not optimal in terms of the convergence rate. Multiscale k-NN and local radial regression are proposed, which can be computed from only the radial distance. Their optimality is also shown.
