SDOclust: Clustering with Sparse Data Observers

Abstract

Sparse Data Observers (SDO) is an unsupervised learning approach developed to cover the need for fast, highly interpretable and intuitively parameterizable anomaly detection. We present SDOclust, an extension that performs clustering while preserving the simplicity and applicability of the original approach. In a nutshell, SDOclust considers observers as graph nodes and applies local thresholding to divide the obtained graph into clusters; later on, observers' labels are propagated to data points following the observation principle. We tested SDOclust with multiple datasets for clustering evaluation by using no input parameters (default or self-tuned) and nevertheless obtaining outstanding performances. SDOclust is a powerful option when statistical estimates are representative and feature spaces conform distance-based analysis. Its main characteristics are: lightweight, intuitive, self-adjusted, noise-resistant, able to extract non-convex clusters, and built on robust parameters and interpretable models. Feasibility and rapid integration into real-world applications are the core goals behind the design of SDOclust.