| This dissertation proposed a boundary-based classification approach, LSVM, which estimates the nonlinear classification boundary using segments of hypeplanes. An innovative vertical neighborhood search technique, EIN-ring approach, using data-mining-ready structure, P-tree, was devised. The strategy of combination of the majority voting approach and boundary-based approach was employed. A comprehensive vertical sample-based KNN/LSVM classification approach with high accuracy and efficiency was proposed where the accuracy mainly results from the combination of the majority voting approach and local boundary approach, and the efficiency is due to the vertical structure, P-tree, and optimized EIN-ring formulations. An interactive KNN/LSVM classification approach was implemented using a server/client scheme that can be extended to a high performance parallel distributed computing system. An extension of vertical EIN-ring KNN/LSVM to density clustering was also proposed. Applications of KNN/LSVM approach and its variations to several real-world application domains, including microarray gene expression data, spatial remote sensor image data, and biological text documents, were analyzed. Finally, a vertical graph modeling for protein and gene interaction network mining was explored for future research work. |
