Visualization of Class Structures using Piecewise Linear Classifiers

摘要

Piecewise linear classifiers are utilized in order to visualize class structures in high-dimensional feature space. The feature space is split into many polyhedral regions by the component hyperplanes of the classifier, and the class label of each polyhedron is determined by taking majority rule on the number of samples. The intra-and inter-class relationships of the polyhedra are shown as a graph, drawing each polyhedron as a node and connecting the nodes by edges according to the adjacency of the regions. Compared to another graph representation method, the proposed method is superior in showing the inseparability or the closeness between the classes.