Textbook of AI

Abstract. Introduces HDBSCAN, the hierarchical extension of DBSCAN that removes the global $\varepsilon$ density-radius parameter. The method builds a hierarchy of clusters across all density levels via mutual reachability distances, then extracts the most stable clusters using a notion of excess of mass. HDBSCAN handles clusters of varying density without per-cluster tuning and produces a soft hierarchy that supports outlier scoring. It is the de facto density-based clustering algorithm in the modern Python ecosystem (hdbscan package).