稳定特征选择的多目标蚁群优化

为了提高进化算法特征选择稳定性，提出一种面向稳定特征选择的多目标蚁群优化方法。通过抽样策略集成三种特征排序法的输出作为多目标蚁群优化的稳定性指导信息，聚合特征的费舍尔分值和最大信息系数值作为多目标蚁群优化的启发式信息，以分类正确率和扩展昆彻瓦指标值作为两个优化目标，兼顾算法的分类性能与特征选择稳定性。在四个标准数据集上进行对比实验，结果表明，所提方法能够在分类性能与稳定性方面达到较好的平衡。
To improve the feature selection stability of evolutionary algorithms, a new method for stable feature selection based on multiobjective ant colony optimization was developed. Feature selection results of three feature ranking methods by resampling policy were combined to provide stable features′ information for multiobjective ant colony optimization; the feature′s Fisher discriminant value and maximal information coefficient value were integrated as heuristic information; the classification correctness rate and value of extensions of Kuncheva similarity measure were taken as two optimization objectives to balance algorithm′s classification performance and its stability. Some comparisons and experiments were carried out on four benchmark data sets, and results show that the proposed method has a better tradeoff between classification performance and feature selection stability.