Clustering Model Selection
🎯 Purpose
Use this card to choose the most appropriate clustering algorithm — KMeans, DBSCAN, HDBSCAN, or Gaussian Mixture Models (GMM) — based on your dataset’s structure, goals, and interpretability needs.
📦 1. When to Use Each Model
| Scenario |
Best Model |
You want fast, scalable clustering with k known |
✅ KMeans |
| You want automatic outlier detection and non-spherical clusters |
✅ DBSCAN |
| You want DBSCAN but with better performance on variable densities |
✅ HDBSCAN |
| You want soft clustering with probability estimates |
✅ GMM |
🧪 2. Model Assumptions & Strengths
| Model |
Shape Assumed |
Strengths |
| KMeans |
Spherical, equal size |
Fast, simple, scalable |
| DBSCAN |
Arbitrary shape, same density |
Detects outliers, no k needed |
| HDBSCAN |
Arbitrary, variable density |
Better noise handling, soft labels |
| GMM |
Elliptical (Gaussian) |
Soft clustering, density modeling |
⚠️ 3. When to Avoid
| Situation |
Avoid... |
| High-dimensional sparse data |
DBSCAN (slow, unreliable) |
| Need explainability |
GMM (less intuitive) |
You don’t know k and can’t guess |
KMeans, GMM (require predefined clusters) |
| Clusters are non-Gaussian or non-convex |
GMM (biased results) |
✅ Decision Checklist
- [ ] Are clusters expected to be well-separated and spherical? → Try KMeans
- [ ] Do you expect noise or arbitrary shapes? → Try DBSCAN or HDBSCAN
- [ ] Is cluster density varied across space? → Prefer HDBSCAN
- [ ] Do you need probabilistic (soft) clustering? → Use GMM
- [ ] Will you tune
k? → Use KMeans or GMM
💡 Tip
“Start with KMeans for speed. Switch to DBSCAN or HDBSCAN for shape. Reach for GMM when you want probabilities — or elegance.”