A Novel Multivariate Feature Ranking Method for Incomplete Categorical Data Based on Weighted Frequency

Feature selection is necessary for high-dimensional data analysis, and several techniques have been proposed after extensive research. Unfortunately, most conventional methods fail when applied to incomplete data, particularly incomplete multi-categorical data. Complete data analysis is a common approach to deal with this problem by imputation and omission, which may introduce immeasurable bias to analysis and requires sophisticated computation. This paper presents a novel feature selection method that combines the weighted frequency with a recursive adaptive partitioning-based MI estimator to rank incomplete multi-categorical features. The proposed method is compared with three other methods with different imputation engines on simulated and real-world data and proved not only improves accuracy but also cuts down the computational time.