Directional density-based clustering

TL;DR

This paper develops a density-based clustering method tailored for directional data on the unit hypersphere, addressing computational challenges and providing tools for parameter analysis, with demonstrated effectiveness through simulations and real data application.

Contribution

It introduces a novel directional density-based clustering methodology for high-dimensional hyperspheres, filling a gap in statistical clustering techniques for directional data.

Findings

Effective clustering on circle and sphere demonstrated through simulations.

Provides a new exploratory tool for smoothing parameter effects.

Successfully applied to exoplanets dataset analysis.

Abstract

Density-based clustering methodology has been widely considered in the statistical literature for classifying Euclidean observations. However, this approach has not been contemplated for directional data yet. In this work, directional density-based clustering methodology is fully established for the unit hypersphere by solving the computational problems associated to high dimensional spaces. We also provide a circular and spherical exploratory tool for studying the effect of the smoothing parameter when kernel density estimation methods are considered. An extensive simulation study shows the performance of the resulting classification procedure for the circle and for the sphere. The methodology is also applied to analyse an exoplanets dataset.