MULTI-CLASS CLASSIFICATION VIA SUBSPACE MODELING

Abstract

Aiming to build a satisfactory supervised classifier, this paper proposes a Multi-class Subspace Modeling (MSM) classification framework. The framework consists of three parts, namely Principal Component Classifier Training Array, Principal Component Classifier Testing Array, and Label Coordinator. The role of Principal Component Classifier Training Array is to get a set of optimized parameters and principal components from each subspace-based training classifier and pass them to the corresponding subspace-based testing classifier in Principal Component Classifier Testing Array. In each subspace-based training classifier, the instances are projected from the original space into the principal component (PC) subspace, where a PC selection method is developed and applied to construct the PC subspace. In Principal Component Classifier Testing Array, each subspace-based testing classifier will utilize the parameters and PCs from its corresponding subspace-based training classifier to determine whether to assign its class label to the instances. Since one instance may be assigned zero or more than one label by the Principal Component Classifier Testing Array, the Label Coordinator is designed to coordinate the final class label of an instance according to its Attaching Proportion (AP) values towards multiple classes. To evaluate the classification accuracy, 10 rounds of 3-fold cross-validation are conducted and many popular classification algorithms (like SVM, Decision Trees, Multi-layer Perceptron, Logistic, etc.) are served as comparative peers. Experimental results show that our proposed MSM classification framework outperforms those compared classifiers in 10 data sets, among which 8 of them hold a confidence level of significance higher than 99.5%. In addition, our framework shows its ability of handling imbalanced data set. Finally, a demo is built to display the accuracy and detailed information of the classification.