Imbalance Data Classification Based on Belief Function Theory

Jiawei Niu,Zhunga Liu

DOI: https://doi.org/10.1007/978-3-030-88601-1_10

2014-01-01

Abstract:Modeling and managing uncertainty in the classification problem remains an important and interesting research topic. Credal classification of uncertain data based on belief function theory has been studied in this thesis, and it allows the object to belong not only to the single classes, but also to any set of classes (called meta-class) with different masses of belief. The credal classification is then of interest to explore the imprecision of classes. Classification methods can be mainly identified by supervised, unsupervised and semi-supervised ones according to the availability of training information. We focus on the supervised and unsupervised classifications. When there are a lot of training samples available in the classification, two credal classifiers for uncertain data are proposed for dealing with different cases. A belief c × K neighbors (BCKN) classifier has been proposed based on belief function theory. In BCKN, the query object is classified according to its K nearest neighbors in each class, and c × K basic belief assignments (BBA?s) are determined according to the distances between the object and these neighbors, and the global fusion of them is used for the credal classification of object. When each class of data can be represented by the prototype vector, a simple credal classification rule (CCR) has been developed using belief functions. Moreover, the missing attribute data is often encountered in classification problem. The different estimations of the missing values can lead to distinct classification results sometimes, and this yields high imprecision and uncertainty of classification due to the lack of information in the missing values.