Performance of the search heuristic

The proposed search heuristic was tested on several data sets. We compared how quickly can VizRank find top ranked projections when using the search heuristic and when evaluating projections in some random order. Below are the results for four data sets - yeast data set and three data sets from UCI repository. We used radviz visualization method and evaluated projections with five of less attributes.

The graphs below show how the score of top ranked projection increases as we increase the number of evaluated projections. In all experiments VizRank was able to find better projections faster when it used the heuristic. We can, however, conclude that the heuristic is most beneficial when analysing data sets with a large number of attributes.

Yeast data set (79 attributes)

Imports-85 (discretized class) data set (25 attributes)

Mushroom data set (22 attributes)

Wine data set (13 attributes)