Density estimation from local structure

Abstract

The authors propose a hyper-ellipsoid clustering algorithm that grows clusters from local structures in a dataset and estimates the underlying geometrical structure of data with a set of hyper-ellipsoids. The clusters are used to estimate a Gaussian Mixture Model (GMM) density function of the data and the log-likelihood scores are compared to the scores of a GMM trained with the expectation maximization (EM) algorithm on 5 real-world classification datasets (from the UCI collection). They show that their approach gives better generalization performance on unseen test sets for 4 of the 5 datasets considered.