K-Means vs DBSCAN: which for geospatial hotspots?

WHAT THIS TESTS: Whether you understand the inductive bias of clustering algorithms and can map them to business context. Interviewers want to see that you do not treat methods as interchangeable black boxes but instead weigh shape assumptions, noise handling, and parameterization burden against the data generation process.

A GOOD ANSWER COVERS: First, the core dichotomy. K-Means is a centroid-based parametric method that partitions the entire dataset into exactly K convex regions by minimizing within-cluster variance. It assumes clusters are isotropic, roughly equal in size, and that every point must belong to one of them. Second, DBSCAN is a density-based non-parametric method that grows clusters from core points using an epsilon neighborhood and a minimum points threshold. It discovers clusters of arbitrary shape, automatically determines the number of clusters, and labels isolated or low-density points as noise rather than forcing membership. Third, the geospatial mapping. Customer locations form irregular anisotropic density peaks around malls or business districts separated by low-density residential or rural zones. DBSCAN naturally captures these irregular hotspots and discards sparse noise, whereas K-Means would slice space into Voronoi cells and assign every house or empty field to a commercial cluster, distorting the analysis.

COMMON WRONG ANSWERS: Picking K-Means for hotspots without mentioning its spherical bias or its inability to leave points unclustered. Claiming DBSCAN requires the number of clusters in advance. Saying K-Means handles outliers well. Suggesting that computational cost alone drives the decision without discussing model assumptions.

LIKELY FOLLOW-UPS: How would you set epsilon and minPoints in DBSCAN for lat-long data? What would you do if hotspot densities varied wildly across cities? How would you handle the fact that distance on a sphere differs from Euclidean distance? Could hierarchical clustering or HDBSCAN improve on DBSCAN here?

ONE CONCRETE EXAMPLE: Suppose you have fifty thousand GPS pings from food-delivery customers in a metro area. K-Means with K equal to five might place a centroid over a river and assign waterfront neighborhoods to a downtown cluster. DBSCAN with a five-hundred-meter epsilon and a minimum of twenty points would instead outline the actual dense commercial corridors, leave parks and industrial zones unlabeled, and reveal a new suburban hotspot you had not expected.