Feature selection technology based on sample unbiased evaluation

This paper focuses on the feature selection methods for unbalanced data sets which have variant sizes of classes. ReliefF has proved to be a successful method for selecting irrelevant features, whereas it is considered as a biased approach for the unbalanced data sets. This paper describes an effective fair method to overcome the defect. Furthermore, against the sensitivity of ReliefF to noisy or irrelevant features when selecting k nearest samples, feature distance is proposed to substitute for the Euclidean distance. Experiments on manual data and UCI data sets indicated that the improved method works better than ReliefF and InfoGain when used as a preprocessing step for naive Bayes and C4.5.