gSVMT: Aggregating SVMs over a dynamic grid learned from data

Addressing the problem of adaptively modelling a classifier as a modular system, a new type of SVM aggregating method termed gridding SVM tree (gSVMT) is proposed in this paper. The proposed gSVMT achieves to discover data subregions with principal discriminant knowledge through a recursive SVM-supervised data partitioning procedure. For each subregion, an individual SVM is allocated to extract the subregion knowledge. A set of such SVMs are aggregated in a specific order, resulting in a globally reliable decision rule to predict new coming samples. Experiments on a synthetic Gaussian data set and 13 benchmark machine learning data sets, have highlighted the usability of the gSVMT on its competitive classification capability. In particular, the proposed gSVMT is found to have better generalization performance than SVM classifiers for data sets with high sparseness and/or class-imbalance. Its performance has been further demonstrated with the successful real application on a face membership authentication system.