靶向量引导的稀疏多目标特征选择算法

Sparse Multi-objective Feature Selection Algorithm Based on Target Vector Guiding Strategy

考虑到滤波法和包装法在特征统计特性和模型评估方面具有互补性优势,而现有混合特征选择算法在高维特征选择问题中求解效率不高.提出了靶向量引导的稀疏多目标特征选择算法(WF-MOFS),充分结合两种方法的优势,并由3个改进策略提高求解效率:(1)基于特征互信息的种群初始化策略,在初代形成良好的前沿面,提高算法收敛速度;(2)设计特征靶向量初始化和更新方式,初始过程考虑特征统计特性,更新过程评估交互特性,靶值越小,特征被选择的概率越高,引导算法进行特征子集的选择;(3)提出稀疏翻转和修复算子,对特征快速降维,并由靶向量指导后代产生高质量特征子集.在15个常用数据集上进行策略有效性、参数选择和算法对比实验,结果表明WF-MOFS算法的有效性和优势.

Considering the complementary advantages of filter method and bagging method in feature statistical characteristics and model evaluation,the existing hybrid feature selection algorithms are not efficient in solving high-dimensional feature selection problems.A sparse multi-objective feature selection algorithm based on target vector guiding strategy(WF-MOFS)is proposed,which fully combines the advantages of the two methods and improves the solving efficiency by three improved strategies:(1)The population initialization strategy based on mutual information forms a good PF in the initial generation and improve the convergence speed;(2)The initialization and update method of feature target vector is designed.The initial target vector considers the statistical characteristics of features,and the update process evaluates their interactive characteristics.The smaller the target value,the higher the probability of feature is selected,and guide the algorithm to select the features;(3)A sparse flip and repair operators are proposed to fast dimensionality reduction for features and generate high-quality offspring under the guidance of target vector.On 15 common data sets,the effectiveness and advantages of WF-MOFS are verified by testing the strategy effectiveness,parameter selection and comparative experiments.