Data Complexity in Pattern recognition, Series: Advanced Information and Knowledge Processing pages:241-248
We use several data complexity measures to explain the differences in classification accuracy using various sets of features selected from samples of magnetic resonance
spectra for two-class discrimination. Results suggest that for this typical problem with sparse samples in a high-dimensional space, even robust classifiers like random decision forests can benefit from sophisticated feature selection procedures, and the improvement can be explained
by the more favorable characteristics in the class geometry given by the resultant feature sets.