Clustering Evaluation Metrics

Silhouette Method

• It measures how well each data point fits within its cluster compared to other clusters
• It calculates a score for each point based on cohesion (how close it is to points in its own cluster) and separation (how far it is from points in the nearest other cluster)
• Score range: −1 to 1
  • 1 = well-clustered
  • 0 = borderline
  • < 0 = likely in the wrong cluster
• applicable to any clustering algorithm that produces cluster labels and works with distance metrics
• commonly used to:
  • evaluate overall clustering quality
  • help choose the optimal number of clusters

Davies–Bouldin Index

• evaluates similarity of clusters based on compactness and separation
• lower is better

Calinski–Harabasz Index

• ratio of between-cluster to within-cluster dispersion
• higher is better

Dunn Index

• higher values indicate better clustering by maximizing inter-cluster distance and minimizing intra-cluster distance