I. Introduction
Dependence and importance analyses are crucial for attribute appraisement and model reconstruction. The corresponding measures in rough set data analysis (RSDA), i.e., dependence degree (DD) and importance degree (ID), have been demonstrated to be useful for successfully solving a variety of problems [1]–[3]. However, DD and ID do not show good reliability and robustness because their values rely on the lower approximation set, which is sensitive to subtle changes of the indiscernibility relation induced by the data. For example, discretizing continuous attributes by different strategies and noises in real data may lead to the changes of indiscernibility relations, which change the values of DD and ID directly. The varying values of DD and ID, which may have big differences, inevitably cause confusion in the process of attribute appraisement and correlation analysis.