A Modified Decomposition Based Multi-objective Optimization Algorithm for High Dimensional Feature Selection

Abstract

Feature selection (FS) is an important research topic in the field of data preprocessing. For this reason, a modified decomposition based multi-objective optimization algorithm, namely M-MOEA/D, is proposed for high dimensional FS, in which an efficient elimination and repair strategy and a modified binary differential evolution (DE) operator are implemented in the decomposition-based framework. Specifically, the elimination and repair strategy is designed based on the symmetric uncertainty. In order to increase the global search capability of the algorithm, a modified binary DE operator is further proposed to cooperate with the elimination and repair strategy. Finally, six different real-world high dimensional data sets are adopted in experiment. The experimental results have validated that M-MOEA/D greatly reduced the size of features set to be selected, and our accuracy was also very competitive when compared to other FS algorithms.