Optimal Classifier Parameter Status Selection Based on Bayes Boundary-ness for Multi-ProtoType and Multi-Layer Perceptron Classifiers

Abstract

Recently, we proposed a new method to select an optimal classifier parameter status (value) using our new criterion that is referred to as uncertainty measure and directly evaluates the Bayes boundary-ness of estimated boundaries. The utility of the method was shown in a task of selecting an optimal Gaussian kernel width, which closely approximates the linear Bayes boundary in the feature space produced by Support Vector Machine classifier. In this paper, we apply the method to two types of classifiers whose class boundaries are basically nonlinear: Multi-ProtoType (MPT) classifier and Multi-Layer Perceptron (MLP) classifier. From experiments using a synthetic dataset and four real-life datasets, we show that our method can provide an optimal (in size and value) classifier parameter status, which basically corresponds to the nonlinear Bayes boundary in given feature spaces, for MPT and MLP classifiers.

This work was supported in part by JSPS KAKENHI No. 18H03266 and MEXT’s Program for Strategic Research Foundation at Private Universities (2014–2018), called “Driver-in-the-Loop”.