Performing classification using all kinds of distances as evidences

Abstract

The classifiers based on the theory of evidence appear well founded theoretically, however, they have still difficulties to nicely deal with the sparse, the noisy, and the imbalance problems. This paper presents a new general framework to create evidences by defining many kinds of distances between the query and its multiple neighborhoods as the evidences. Particularly, it applies the relative transformation to define the distances. Within the framework, a new classifier called relative evidential classification (REC) is designed, which takes all distances as evidences and combines them using the Dempster'rule of combination. The classifier assigns the class label to the query based on the combined belief. The novel work of this method lies in that a new general framework to create evidences and a new approach to define the distances in the relative space as evidences are presented. Experimental results suggest that the proposed approach often gives the better results in classification.