We propose new classification methods based on successive separation and boosting multi-hyperplane separation. The successive separation procedure is based on a linear programming formulation, whose effectiveness is achieved by constituting an iterative tree structure to yield multiple sub-regions for separating points of two groups, and refined by retrospective enhancement. The boosting multi-hyperlane separation is an alternative to the successive procedure. It combines different "weak" classifiers into a single accurate classifier. Instead of using weak base classifiers, we take advantage of a linear programming discriminant model and achieve better separation performance.;We focus on two-class classification, but also extend our approaches to multi-class separation problems. We explore model implications in a computational study of several important data mining applications, and undertake computational and analytical comparisons with three main hyperplane-based classification techniques, including Oblique Decision Trees, Piecewise Linear Models and Support Vector Machine. Our proposed methods offer significant promise for improving the accuracy and efficiency of classification. |