📏Silhouette Score

NEW

Also known as: Silhouette coefficient

A clustering validation metric measuring cohesion and separation by comparing intra-cluster tightness to nearest-cluster distance.

DESCRIPTION

The silhouette score evaluates clustering quality by computing, for each sample, its average distance to points in its own cluster and the nearest neighboring cluster. Values range from −1 to 1, with higher indicating better-defined clusters. It is model-agnostic and helpful for selecting the number of clusters or comparing algorithms. However, it depends on the chosen distance metric and can mislead on complex shapes or varying densities. Use alongside other indices and domain checks for robust assessment.

ALTERNATIVE DEFINITIONS

Wikipedia

The silhouette value is a measure of how similar an object is to its own cluster compared to other clusters.

Defines the sample-level interpretation central to the score.

Source

scikit-learn

The Silhouette Coefficient is defined for each sample and is composed of two scores: a, the mean intra-cluster distance, and b, the mean nearest-cluster distance.

Clarifies the a and b components that determine sample coefficients.

Source

RELATED TERMS

Cluster Analysis K-Means K-Medoids Dissimilarity Measure Intra-cluster Distance Inter-cluster Distance Unsupervised Learning Cluster Validation

COMPETING TERMS

Davies–Bouldin Index Calinski–Harabasz Index

RELATIONS

• Broader: Score

KNOWLEDGE AREAS

Data Science AI Mathematics Testing & Validation

TARGET AUDIENCE

Data Scientist Data Analyst AI Engineer

EXTERNAL SOURCES

• Wikipedia - Silhouette (clustering)
• scikit-learn - Documentation on silhouette coefficient
• R cluster package - Silhouette implementation in R

TOP QUESTIONS

When should I use the Silhouette Score?

Use it to compare clustering results and estimate a suitable number of clusters when clusters are globular and separable under your chosen distance metric.

Category: Application

What is a good Silhouette Score?

Scores near 1 indicate well-separated clusters, around 0 suggest overlapping clusters, and negative values indicate potential misassignment.

Category: Interpretation

What are common mistakes with Silhouette Score?

Relying on it alone can mislead with non-spherical clusters or varying densities. Always pair with alternative indices and visual inspection.

Category: Mistakes

How does distance choice affect Silhouette Score?

Different metrics (e.g., Euclidean vs. cosine) alter intra- and inter-cluster distances, changing the score and conclusions.

Category: Concept

Can Silhouette Score select k in k-means?

Yes, compute scores across k values and choose the maximum, but validate with domain knowledge and other metrics.

Category: BestPractices

Does Silhouette Score work for density-based clustering?

It can be applied, but performance may be suboptimal if cluster shapes defy metric assumptions. Consider DBSCAN-specific diagnostics too.

Category: Alternatives

How is the Silhouette Score aggregated?

It averages sample-level coefficients to a global score, optionally computing per-cluster averages for detailed diagnostics.

Category: Basics

Is the Silhouette Score scale-invariant?

No. Feature scaling affects distances and thus the score; standardize features when appropriate.

Category: Requirements

What is a quick diagnostic alongside Silhouette?

Plot silhouette bars per cluster and a 2D embedding (e.g., PCA or t-SNE) to visualize separability and cluster compactness.

Category: Practice