Clustering
Clustering groups unlabeled points by similarity; the method and distance you pick define what a cluster even means.
Finding groups without labels
Clustering partitions unlabeled data into groups whose members are more similar to each other than to those in other groups. It is a core unsupervised task used for exploration, segmentation, and compression. Crucially, there is no single correct clustering; the answer depends on the algorithm, the distance metric, and what 'similar' means for the problem.
Families of methods
- Centroid-based (k-means): partitions into k spherical clusters around means; needs k.
- Density-based (DBSCAN): clusters as dense regions, finds arbitrary shapes and outliers.
- Hierarchical: builds a tree of nested clusters, cut at any level.
- Distribution-based (Gaussian mixtures): models clusters as probability distributions, soft assignments.
See k-means, DBSCAN, hierarchical clustering, and Gaussian mixtures for each.
Evaluating clusters
- Internal (no labels): silhouette score, Davies-Bouldin index, within-cluster inertia.
- External (labels available): adjusted Rand index, normalized mutual information.
- Stability: does the clustering hold up under resampling and reseeding.
Practical guidance
Scale features first, since most methods use distances. Pick the algorithm by the cluster shape you expect: k-means for compact blobs, DBSCAN for irregular shapes with noise, hierarchical when you want multiple granularities. Validate that clusters mean something to the domain, not just that a metric improved. In fusion diagnostics, clustering raw operating data can surface distinct plasma regimes that no one labeled in advance, guiding later supervised analysis.