A novel missing-rate-oriented selective algorithm for handling missing data by minimizing imputation

Abstract

A novel algorithm named Missing-Rate- Oriented Selective (MROS) algorithm - including: Most-Similar (M-S) algorithm and Attribute-Selective Imputation (ASI) approach- is proposed to achieve effective Mean Identification Rate (MIR) with minimal imputation effort for multi-classification systems in a complex and High Missing Rate (HMR) dataset. This dataset was developed from real server power supply failure cases which are characterized by categorical variables and 96.66% of the samples contain missing values - with 43.33% of the samples in the HMR region (40%~64.44%). Experiments prove MROS algorithm is capable in achieving effective MIR over the full missing rate range with minimal imputation effort, notably achieving 80%~86% MIR in HMR region with an imputation rate between 15.61% and 25.07%.