A New Belief-based Classification Fusion for Incomplete Data

Abstract

Reducing the negative impact of estimation on classifier performance in training set is one of the most challenging tasks in incomplete data classification. A new belief-based classification fusion method (BCF) is proposed for incomplete data in this paper and the core idea is to make full use of the existing attributes of incomplete objects in training set to improve the performance of basic classifier without deleting or estimation strategy. Specifically, for a data set with n-dimensional attributes, different attributes generate p (p ≤ n) subsets according to prior knowledge or random combination. Then, $p$ trained basic classifiers (such as SVM) will be obtained with complete objects from corresponding $p$ training subsets, and estimation strategy is used to fill the incomplete objects in the test set. Finally, DS rule is used to fuse $p$ sub-classification results if they do not conflict and a new global fusion method is proposed to fuse the remaining conflict sub-classification results, which can submit the object difficult to be accurately classified into a singleton (special) class to meta-class to reduce error rate and characterize the uncertainly caused by missing values well. Our simulation results illustrate the potential of the proposed method using real data sets, and they show that BCF can improve substantially the classification accuracy.