Feature space vs embedded space

The distinction

Ounias supports clustering in two different analytical spaces.

Feature-space clustering operates directly on the selected analytical features of the dataset.

Embedded-space clustering first generates a numerical lower-dimensional representation of those features and then performs clustering on the resulting coordinates.

Feature-space clustering

The clustering algorithm receives the selected dataset features directly.

• No dimensionality-reduction step is required
• The resulting labels depend directly on the selected feature representation
• Distances and structure are determined in the original analytical feature space

Embedded-space clustering

An embedding is generated first, and the clustering method operates on those coordinates.

The embedding becomes part of the analytical configuration. Changing the embedding method or its parameters can therefore change the structure presented to the clustering algorithm.

• Useful for lower-dimensional representations
• Allows nonlinear representations to be used before clustering
• Embedding and clustering choices should be interpreted together
• The embedding is preserved as part of the run

How the two approaches relate

Neither analytical space is automatically the correct one.

Feature-space and embedded-space clustering are alternative representations of the same dataset rather than interchangeable implementations of the same analysis.

Ounias keeps them separate so their results can be evaluated and compared rather than silently treating one representation as authoritative.