Description
Evaluates whether user segments from cluster analysis are real, stable and useful using internal indices, resampling stability and external validation, not just visual inspection.
Cluster algorithms always return clusters, whether or not the data contain meaningful groups. Cluster validation tests whether a segmentation reflects real structure. Internal validity uses indices such as the silhouette coefficient, Calinski-Harabasz and Davies-Bouldin to compare solutions; the gap statistic compares clustering against a null reference. Stability is assessed by bootstrapping or subsampling and measuring agreement between solutions (e.g. Jaccard coefficients or adjusted Rand index). External validity checks whether clusters differ on variables not used to create them, such as retention or revenue, and whether they make sense qualitatively. Rigour comes from pre-specifying the candidate range of clusters, standardising variables appropriately, and reporting weak structure honestly.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Definition
- Data
- Quantitative
- Effort
- Days
Steps
- Cluster validation protocol:
- Run clustering for k = 2..10.
- (2) For each k, compute: silhouette score (higher is better, range -1 to 1; good clusters > 0.5), Calinski-Harabasz index (higher is better), Davies-Bouldin index (lower is better).
- (3) Plot all three indices vs. k. Identify the k where the silhouette score peaks or the 'elbow' in CH/DB indices.
- (4) Also compute gap statistic (compares within-cluster dispersion to random data).
- (5) Report: which k did you choose, and why? Include all validation metrics.
Sources
- Cluster Analysis for Researchers — (various)
Also known as
- Cluster validity assessment
- Silhouette analysis
- Segmentation validation
Categories
- Analytics & Behaviour
- Statistics & Modelling