On an Improvement to Initial Value Dependency Problem of Two Fuzzy Clustering Algorithms for Categorical Multivariate Data

Abstract

The results from fuzzy clustering algorithms are heavily affected from initial values, and an easy initial value setting produces local optimal solution or saddle point, resulting in poor clustering accuracy. In this report, an initial value setting is proposed using eigen pairs of Hessian for the objective function. Then, this setting is applied to the entropy-regularized and q-divergence-based fuzzy clustering algorithms for categorical multivariate data.