B1452
Title: Mixtures of locally-mapped support vector machines
Authors: Hien Nguyen - La Trobe University (Australia) [presenting]
Geoffrey McLachlan - University of Queensland (Australia)
Abstract: Support vector machines (SVMs) have been highly successfully in application to classification problems of all sizes. However, the usual construction of SVMs only allow for a singular mapping between input and classification, even if that mapping is nonlinear, such as in the case of kernel SVMs. Via recent approaches to local-mapping of mixture regressions, we construct an approach to SVM classification that allows for different mappings in different parts of the input domain. We demonstrate how our mixtures of locally-mapped SVMs can be estimated via maximum quasi-likelihood estimation and the MM (minorization-maximization) algorithm. Furthermore, we present an online algorithm for its estimation in the face of big data. Some theory regarding the estimator and algorithms are presented, and example applications of the methodology are provided.
