An approach of feature space evaluation based on evidence theory

Feature selection and extraction is one of the most important and essential problems in pattern classification. The basic task of feature selection and extraction is how to obtain the most useful and important features by selection or transformation, so the evaluation of feature spaces is needed. The traditional feature space evaluation approaches are always based on the discernibility measures directly defined over the feature spaces of samples. A new feature space evaluation approach is proposed. The original feature spaces of samples are first transformed to the evidential spaces. Then by using distance of evidence, the evidential discernibility measure is defined to indirectly describe the discernibility of the original feature spaces. Experimental results show the rationality and efficiency of the proposed approach.