Randomization Test for Importance Degree of Variables in Rough Set Theory

Abstract

The appraisement of variable importance and contribution is the central problem for variable selection and relevance analysis, particularly in the domains of ecological and medical science. Except for statistical modelling, more interesting methods, such as rough set and artificial neural network, are used to analyze the variable contribution in systems. But the results derived from rough set and statistical theory can not be compared with each other because the lack of common descriptions. In this paper, we propose and demonstrate a randomization test for statistically assessing the variable importance degree in rough set theory. The randomization approach can identify variables that significantly contribute to the predictions of the system and reach a more objective result which can be compared with statistical analysis. Thus, the bridge of rough set theory and statistical approach is constructed. Furthermore, by the randomization test, the interaction of variables can be easily appraised and the variables which have the same importance degrees can be distinguished. At last, an experiment shows the function of randomization test for importance degree of variables in rough set theory.