Spectral Clustering

Uses eigenvalues of similarity matrix to perform dimensionality reduction before clustering in fewer dimensions.

When to use:

• Have non-convex clusters (circles, spirals)
• Graph-structured data
• Small to medium datasets
• Need to capture complex cluster shapes

Strengths: Handles non-convex shapes, works with graph data, can capture complex patterns, no assumptions about cluster shape Weaknesses: Slow on large datasets, requires specifying k, sensitive to parameters, memory intensive

Model Parameters

N Clusters (default: 8, required) Number of clusters to form.

Affinity (default: "rbf") How to construct the similarity matrix:

• rbf: Radial basis function kernel (default, good for continuous data)
• nearest_neighbors: K-nearest neighbors graph
• precomputed: Use your own affinity matrix

N Neighbors (default: 10) Number of neighbors for nearest_neighbors affinity.

• 5-10: Local structure
• 10-20: Good default
• 20+: More global structure

Assign Labels (default: "kmeans") Strategy for assigning labels in embedding space:

• kmeans: Use K-Means (default, usually better)
• discretize: Discretization method (faster)

Random State (default: 42) Seed for reproducibility.