When observations have no predefined labels, clustering seeks structure by grouping similar observations. Unlike supervised methods, it has no known outcome to predict; its groupings can instead support later analysis and decisions. section 5.3.4
Two common approaches construct those groupings differently:
- K-means assigns observations to their nearest centroid, then iteratively updates centroids until assignments stabilize. section 5.3.4
- Hierarchical clustering builds a dendrogram of nested groupings, either by successive merging or successive splitting. section 5.3.4
Both approaches group observations through geometric relationships in feature space, aiming for similar members within clusters and distinctions between clusters. Geometry does not directly measure how much uncertainty a grouping removes; information theory reframes evaluation around that question. section 5.3.4 section 5.3
Where clustering can reveal patterns
Applications include grouping customers by purchasing behavior, patients by clinical characteristics, and social-network members into communities. section 5.3.4























