SPORE: Skeleton Propagation Over Recalibrating Expansions

TL;DR

SPORE is a novel clustering algorithm that effectively handles arbitrary cluster geometries by propagating skeletons and recalibrating expansions, outperforming existing methods across diverse datasets.

Contribution

Introduces SPORE, a clustering method that adapts to arbitrary geometries without global density parameters, using skeleton propagation and local distance statistics.

Findings

SPORE significantly outperforms baseline algorithms in ARI recovery across 28 datasets.

Achieves strong performance with only ten hyperparameter evaluations.

Handles nested and asymmetrically separated structures effectively.

Abstract

Clustering is a foundational task in data analysis, yet most algorithms impose rigid assumptions on cluster geometry: centroid-based methods favor convex structures, while density-based approaches break down under variable local density or moderate dimensionality. This paper introduces SPORE (Skeleton Propagation Over Recalibrating Expansions), a classical clustering algorithm built to handle arbitrary geometry without relying on global density parameters. SPORE grows clusters through a nearest-neighbor graph, admitting new points based on each cluster's own evolving distance statistics, with density-ordered seeding enabling recovery of nested and asymmetrically separated structures. A refinement stage exploits initial over-segmentation, propagating high-confidence cluster skeletons outward to resolve ambiguous boundaries in low-contrast regions. Across 28 diverse benchmark datasets, SPORE achieves a statistically significant improvement in ARI-based recovery capacity over all evaluated baselines, with strong performance accessible within ten evaluations of a fixed hyperparameter grid.