一种基于成对样本比较的马田系统特征筛选方法OACHSSCD

The traditional Mahalanobis-Taguchi System(MTS),as a feature screening method,has two shortcomings:Firstly,it relies on samples with clear category labels for training,which requires high prior information;secondly,it uses signal-to-noise ratio as the evaluation function for the feature subset,making it difficult to ensure that the screened features have strong classifica-tion discrimination ability.To this end,the paper proposes a feature screening method for the MTS based on pairwise sample com-parison.This method,on the one hand,collects both similar and non-similar pairs of samples through"pairwise comparison"for training,thereby reducing the reliance on prior information;on the other hand,based on the goal of"bringing closer similar sam-ples and pushing away non-similar samples",a new feature subset evaluation function is constructed to replace the sig-nal-to-noise ratio,and the Taguchi orthogonal experiment process is redesigned.Experimental results show that the performance of the proposed method is significantly superior to the unsupervised methods(Variance,Laplacian Score)on four UCI datasets,and is close to or slightly better than the supervised methods(Relief,Fisher Score).In an example of sudden epidemic screening based on clinical data from a certain city from 2020 to 2023(1099 samples,40 indicators),the feature subset screened by the proposed method under weak supervision can ensure a leakage screening rate of less than 1.5%and a misjudgment rate of less than 10%as the risk prevention bottom line,and significantly reduces the comprehensive risk cost,verifying its effectiveness and practical ap-plication value.