DBSCAN

A core observation has enough neighbors within the specified radius according to the method's minimum-sample rule. Connected core observations form dense regions, and nearby border observations can join those clusters without themselves meeting the core criterion. Remaining points are noise. The cluster count is an outcome rather than an input, and groups need not be spherical. A single radius applies across the data, which can be difficult when densities differ greatly. DBSCAN is distinct from centroid clustering and from HDBSCAN's hierarchy over density levels. Understanding neighborhood geometry is essential because feature scaling can change density and therefore the entire cluster structure.

Select a distance appropriate to the representation and investigate scaling before setting the neighborhood radius. Compare plausible radius and minimum-sample values and inspect core, border and noise behavior. Check sensitivity to sampling and ambiguous observations, especially where groups are close. Estimate neighborhood computation and memory requirements for the intended dataset. The deliverable should describe clusters, unassigned cases and parameter choices, with domain inspection establishing whether dense groups correspond to the question rather than only satisfying the algorithm's connectivity rule.

One radius may not represent clusters with strongly different densities, and high-dimensional distance can become uninformative. Dense bridges can merge groups that an analyst expected to remain separate. Border assignments can depend on implementation details or ordering in ambiguous cases. Noise does not mean invalid data. DBSCAN also does not directly define a general supervised class predictor. Assess parameter sensitivity and substantive usefulness before treating a density partition as a stable taxonomy.