Fig. 2

A comparative evaluation using diceR applied to three datasets. Using 10 clustering algorithms, we repeated the clustering of each data set, each time using only 80% of the data. Four ensemble approaches were considered. The ensembles were constructed using all the individual clusterings and were repeated by omitting the least performing algorithms (the trim version in the figure). Thirteen internal validity indices were used to rank order these algorithms based on performance from top to bottom. Indices were standardized so their performance is relative to each other. The green/red annotation tracks at the top indicate which indices should be maximized or minimized respectively. Ensemble methods were highlighted using a bold font