SmartStripes - Looking under the Hood of Feature Subset Selection Methods

Abstract

We propose a visualization method for the diagnosis and interactive refinement of automatic techniques for feature subset selection. So-called filter techniques use statistical ranking measures to identify the most useful combination of features for further analysis. Usually a measure is applied to all entities of a data-table. The influence of atypical entities can distort the result, but this distortion may be masked by the statistical aggregation. Clearly, feature and entity subset selection are highly interdependent. Our technique, SmartStripes, intends to make this interdependency visible.