Fig. 2.

Clustering quality was assessed using seven different measures for each pair of algorithm and dataset: AMI, ARI, F-measure, homogeneity, majority, silhouette and VI distance. Principal component (PC) analysis with feature scaling was then performed on the collection of 102 clustering partitions in the space defined by the quality measures. The heatmap shows absolute Pearson correlation among the seven different measures, as well as the top three principal components (PCs). The latter collectively explain over 90% of the variance in the measurement data, as indicated in their labels. A group of four different measures: AMI, ARI, F-measure and VI are best correlated with PC1, which captures 61% of the variation. Two other measures, the homogeneity and majority, are also highly correlated and are best reflected by PC2. The remaining silhouette measure is represented by PC3.