seniorK-Means Clustering
Why does K-Means fail on non-convex cluster shapes?
Updated May 16, 2026
Short answer
K-Means fails on non-convex shapes because it assumes clusters are spherical and separable by linear boundaries.
Deep explanation
K-Means partitions space using Voronoi cells around centroids, which are inherently convex. Non-convex structures like spirals or moons cannot be represented by a single centroid without incorrect splits. This leads to fragmentation of true clusters into multiple artificial groups.
Real-world example
Image segmentation where objects have curved or irregular shapes.
Common mistakes
- Forcing K-Means on manifold-structured data instead of using density-based methods.
Follow-up questions
- Which algorithm handles non-convex shapes better?
- Why is centroid representation limiting?