An Incremental Feature Selection Approach Based on Information Entropy for Incomplete Data
Chuan Luo, Tianrui Li, Yi Zhang · 2019
Data uncertainty has become increasingly important owing to the dynamic and incomplete characteristics of data. The selection of nonredundant features from such uncertain data is a highly challenging problem. Information theory can help us to make qualitative measures about the uncertainty for analyzing data, which has been widely used to feature selection task. In this paper, an incremental feature selection approach from information-theoretic perspective for dynamic incomplete data is presented. By utilizing the updating mechanisms of classification induced by conditional and decision features, a novel incremental representation of Shannon's entropy for incomplete data is proposed to accelerate the computation of feature significance in heuristic searching process. Efficient incremental feature selection algorithm is developed when the incomplete data increase dynamically in size. Theoretical justifications and illustrative examples are provided to demonstrate the validity of the proposed algorithm.