Abstract: The current filter feature selection models use greedy strategy combined with mutual information to evaluate feature subsets， which are easy to fall into the local optimum trap. Considering the effect of label information on redundancy， an improved MIFS-U method is used to measure the redundancy under the condition of a given label. A decomposition-based multi-objective optimization framework combined with a differential evolution operator that introduces polynomial mutation is used for global search to avoid searching into local optimum. The l₁ regularization is introduced to ensure the sparsity of the feature subset， and a new feature selection algorithm MOEA/D-DEFS is proposed. In the experimental stage， the knn-5 classifier is used to verify the learning effect， by the tests on multiple sets of data sets from different fields. The results show that considering feature selection as a multi-objective problem and using a global search strategy can provide better performance in terms of feature subset dimensions and classification accuracy.

Key words: feature selection, mutual information, multi-objective evolution algorithm, l₁-regularization, redundancy