Clustering methods

Overview

Clustering groups observations according to patterns in the selected analytical space.

Ounias can cluster the original selected features or a lower-dimensional representation generated by an embedding method.

Choosing an analytical space

Clustering can operate on selected dataset features or on coordinates generated by an embedding method. The choice affects which structure the method receives.

Supported clustering methods

Different methods make different assumptions about group shape, density, and membership.

• Centroid and distribution based — K-Means: Partitions observations around a selected number of cluster centers. [Feature space, Embedded space]
• Centroid and distribution based — Gaussian Mixture Model: Models observations as a mixture of probabilistic distributions. [Soft membership, Feature space, Embedded space]
• Density based — DBSCAN: Finds dense regions while allowing some observations to remain unassigned. [Density based, Noise aware]
• Density based — HDBSCAN: Extends density-based clustering across varying density levels. [Density based, Variable cluster count]
• Structural — Agglomerative clustering: Builds a hierarchical grouping by progressively combining observations or clusters. [Hierarchical, Feature space, Embedded space]
• Structural — Spectral clustering: Uses relationships between observations to identify potentially nonlinear group structure. [Graph based, Nonlinear structure]

Evaluation and enrichment

• Separation and structure metrics
• Standardized metric pillars
• Stability analysis
• Feature explanations
• Rule extraction
• Substructure analysis
• Cross-run comparison
• Longitudinal analysis where applicable

Interpretation and limitations

• Clusters are analytical representations, not automatically meaningful real-world groups.
• Results can depend on the selected features, embedding, distance structure, algorithm, and parameters.
• Domain interpretation remains necessary when assigning meaning to discovered groups.