MFC: Initialization method for multi-label feature selection based on conditional mutual information

초록

Feature selection is widely used in multi-label classification because of its simplicity, efficiency, and accuracy. Specifically, evolutionary algorithm (EA)-based multi-label feature selection methods have produced good potential. However, in conventional EA-based approaches, the initial population is usually generated at random, because knowledge about the input dataset is not usually available. In this study, we propose a method for generating the initial population of an EA-based multi-label feature selection method considering dependencies between features and labels. To the best of our knowledge, this is the first initialization method described for EA-based multi-label feature selection. The initial population generated by the proposed method can be easily applied to the EA-based feature selection method because it is independent of EA-based feature selection algorithms. © 2019

키워드