Clustering
Grouping data points so that members of a group are more similar to each other than to points in other groups.
Definition
Clustering partitions data into groups (clusters) of similar points without using labels. It is a core unsupervised task, answering the question of what natural groupings exist in the data.
Because clustering has no ground truth, results should be treated as hypotheses about structure rather than facts. Running several methods and checking whether the same groupings recur is a practical guard against reading meaning into artifacts of one algorithm's assumptions.
The number of clusters is rarely given by nature, and forcing a partition where none exists produces confident nonsense. Good practice combines quantitative indices with domain knowledge and with a check of whether the clusters are stable under resampling and reproducible across methods. Density-based methods add value by declining to assign outliers, acknowledging that not every point belongs to a group.
Families of methods
- Centroid-based: k-means assigns points to the nearest cluster center.
- Hierarchical: builds a tree of nested clusters.
- Density-based: DBSCAN finds dense regions and labels sparse points as noise.
- Model-based: Gaussian mixtures fit overlapping distributions.
Choosing the number of clusters
Many methods require the number of clusters in advance. Heuristics such as the elbow method, silhouette score, or gap statistic help choose it, but the right number is often a modeling judgment rather than a fact.
Why it matters
Clustering reveals structure, segments populations, and compresses data into representative groups. It is exploratory: results depend on the distance metric and scaling of features.
Fusion connection
Clustering ensembles of Hyperion simulations groups operating points into families with similar plasma behavior, giving engineers a manageable map of a large design space.