82 A Consensual Subspace Method to Enhance Classification Accuracy
-
Published:2008
Download citation file:
The proposed method uses different input features to partition the sample space in to subspaces in a two-level decision treelike structure to enhance the performance of a classifier. The support vector machine is used as the classifier in this paper. Each input feature used is associated with a threshold such that an input vector traverses to either the right node or the left node of a parent node. Given a feature, the best threshold is usually found by minimizing a measure such as the impurity, characterized by Gini index or information entropy. In this way, for each pair of a feature and the corresponding threshold, the data is partitioned in to two groups. Each group is trained with a specialized SVM. During testing, each data point is directed to one of the SVM's based on the feature used and its threshold. The method is further generalized by choosing a subset of rank-ordered features. For this purpose, an impurity measure is used. In this way, a number of subspace classifiers are generated. In the end, the final classification is done by consensus between the subspace classifiers. This usually results in better accuracy as compared to a single SVM classification.