DBSCAN is a density-based clustering algorithm in unsupervised machine learning that defines clusters by local point density rather than by distance to a centroid, allowing it to recover clusters of arbitrary, non-convex shape and to leave low-density points unassigned as outliers. It rests on a chain of formal definitions built from two parameters, ε (the radius of the neighborhood over which points are counted) and minPts (the density threshold): a point is a core point if its ε-ball contains at least minPts points; a point is density reachable if it can be reached from a core point by successive hops of length at most ε passing only through core points; and two points are density connected if some core point renders both density reachable. The cluster is then defined exactly as an equivalence of density connectedness — two points lie in the same cluster if and only if they are density connected — which distinguishes DBSCAN from centroid-based methods such as k-means and from hierarchical clustering, whose cluster shapes are constrained by their distance or linkage criteria.
DBSCAN is a density-based clustering algorithm in unsupervised machine learning that defines clusters by local point density rather than by distance to a centroid, allowing it to recover clusters of …