A Robust Feature Selection Algorithm

The paper proposes a novel feature selection algorithm based on the density function of features. Density difference between features for paired classes is used for assigning weights to the features. Existing methods select only those features that can distinguish between all classes at the same time. A new filter based feature selection method termed as Paired Class Density Difference with Inconsistency (FSDD) is proposed in this paper. It can be used for a multi-class problem. This approach selects those features that distinguish between few classes but still play an important role in classification. Inconsistency measure is used in order to remove redundancy in the selected feature set. Classification accuracy of the proposed method is compared with that obtained using existing filter based feature selection methods on UCI machine learning repository datasets and on manual segmentation data provided by NIST. Increased classification accuracy for FSDD shows that the concept of using density difference to assign feature weight is a significant contribution.