Handling Different Preferences Between Objectives for Multi-objective Feature Selection in Classification

Maximizing the classification performance and minimizing the feature subset size are two key objectives in multi-objective feature selection. Most existing works treat these two objectives equally. However, from the perspective of decision-makers, the preferences of these two objectives are different, that is, the classification performance is more important than the number of selected features. Besides, improving the classification performance is also more challenging than reducing the number of selected features. To deal with this issue, this paper proposes a preference-inspired multi-objective evolutionary algorithm, which consists of three major components: 1) a fitness function is proposed to give more preference to the objective of classification performance; 2) based on the analysis of solutions' distribution, an irrelevance learning method is proposed to detect the irrelevant features; 3) a dimensionality reduction method is proposed to remove irrelevant features and further improve the classification performance of feature subsets. By comparing the proposed method with five state-of-the-art multi-objective evolutionary algorithm-based feature selection methods, empirical results on nine classification datasets demonstrate that the proposed method is able to obtain a set of feature subsets with better classification performance.