Chunking for Massive Nonlinear Kernel Classification

Abstract

A chunking procedure [2] utilized in [18] for linear classifiers is proposed here for nonlinear kernel classification of massive datasets. A highly accurate algorithm based on nonlinear support vector machines that utilizes a linear programming formulation [15] is developed here as a completely unconstrained minimization problem [17]. This approach together with chunking leads to a simple and accurate method for generating nonlinear classifiers for a 250000-point dataset that typically exceeds machine capacity when standard linear programming methods such as CPLEX [12] are used. Because a 1-norm support vector machine underlies the proposed method, the approach together with a reduced support vector machine formulation [13] minimizes the number of kernel functions utilized to generate a simplified nonlinear classifier.