Incomplete data ensemble classification using imputation-revision framework with local spatial neighborhood information

Abstract

Most existing machine learning techniques require complete data. However, incomplete patterns are common in many real-world scenarios due to the missing values (MVs). Various Missing value imputation (MVI) methods have been proposed to recover the MVs. Each of them has its own advantages in some scenarios. However, on the one hand, few of them consider taking advantages of different MVI methods; on the other hand, how to improve the imputation performance with local information is still an open problem. This paper proposes an imputation-revision framework with local spatial neighborhood information for incomplete data classification. The proposed method endeavors to combine the advantages of several imputation methods. It first obtains several complete datasets which are pre-filled by various MVI methods. Then, it detects the local neighborhood information (LNI) of samples and revises MVs based on the LNI. Finally, ensemble technique is employed to give a final decision. Numerical experiments have verified the superiority of the proposed method in terms of both prediction accuracy and algorithm stability.