Non-Myopic Feature Quality Evaluation with (R)ReliefF

Researchers in machine learning, data mining, and statistics have developed a number of methods that estimate the usefulness of a feature for predicting the target variable. The majority of these measures are myopic in a sense that they estimate the quality of one feature independently of the context of other features. Our aim is to show the idea, advantages, and applications of nonmyopic measures, based on the Relief algorithm, which is context sensitive, robust, and can deal with datasets with highly interdependent features. For a more thorough overview of feature quality measures, see [15].