To overcome the blindness of subjective selecting dimensionless indictors of target features as sensitive features
with less experience, a target feature sensitivity evaluation method based on clustering analysis and geometry is proposed.
The method defines “compact degree”, “dissociative degree” and “incompact-compact degree” of samples distributing in
features space, and introduces the knowledge of rotundity from geometry, by which the classing ability of target feature to
multi-class target samples is reviewed. Then the new combined features are constructed through the sensitivity evaluation
results, and the classify recognition efficiency can be improved. Finally, simulation results show the effectiveness of the
method.