Abstract

According to the information theories, this present study has carried out an analysis towards the concept and properties of the cross entropy. On this foundation a concept of symmetry cross entropy (SCE) is put forward and proved that the SCE satisfies three axiomatization requested of the distance. SCE is a distance measure, which is used to measure the degree of variation between two-class problems. Based on the SCE, the average symmetry cross entropy (ASCE) is initialized, and it is used to measure the degree of variation between multi-class problems. Regarding the ASCE separability criterion of the multi-class for information feature compression, a novel information feature compression algorithm is constructed in this present study. The proposed algorithm considers the class information, so it is a supervised algorithm of feature compression. The experimental results demonstrate the algorithm is valid and reliable, and it provides a new approach for information feature compression in pattern recognition.

Paper Details