Dual of nearest-class-model methods: A separating hyperplane classification framework

Speaker:  Rui Zou, University of Kent

Abstract

Nearest-class-model methods are an important classification scheme, which first builds class models separately for each class and then classifies a test sample to the class with the shortest distance from the test sample to the class model. Such a scheme makes adding new classes easy. In this work, we first propose a nearest convex cone method (NCCM) to fill the tightness gap between two existing nearest-class-model methods, the nearest subspace method (NSM) and the nearest convex hull method (NCHM). Then we investigate NSM, NCHM and NCCM both theoretically and empirically, to understand deeply their underlying classification mechanisms and analyse their data-dependent classification performances. The novelties of our work are threefold: we propose NCCM and provide new theoretical dual analysis of NCCM; we establish a novel separating hyperplane classification (SHC) framework to unify the nearest-class-model methods; and we design a new data-exploration scheme to analyse the characteristics of a dataset and explain why such characteristics make one or more of the methods suitable for the data.