An Adaption of Relief for Redundant Feature Elimination

Feature selection is important for many learning problems improving speed and quality. Main approaches include individual evaluation and subset evaluation methods. Individual evaluation methods, such as Relief, are efficient but can not detect redundant features, which limits the applications. A new feature selection algorithm removing both irrelevant and redundant features is proposed based on the basic idea of Relief. For each feature, not only effectiveness is evaluated, but also informativeness is considered according to the performance of other features. Experiments on bench mark datasets show that the new algorithm can removing both irrelevant and redundant features and keep the efficiency like a individual evaluation method.