Variable selection in discriminant analysis based on Gram-Schmidt process

Huiwen Wang (1) , Meiling Chen (1) , Gilbert Saporta (2)

1 BUAA - Beihang University
2 CEDRIC - MSDMA - CEDRIC. Méthodes statistiques de data-mining et apprentissage

A new linear discriminant analysis modeling method based on Gram-Schmidt process was introduced, which firstly selected the most effective variables for classification in the independent variables set. In the meantime, the insignificant variables and the redundant information were identified and removed from the independent variables set. The selected variables were transformed into a set of orthogonal vectors by Gram-Schmidt process. Not only can the proposed method accomplish variable selection in linear discrimination, but also overcome the multi-collinearity problem effectively. Since F-statistic works as a criterion to verify the discrimination effect of each selected variable, it helps analysts to understand the analysis result. In order to test the reasonableness and effectiveness of the method, a simulation experiment was carried out. The result indicates that the proposed method can lead to a reasonable and precise conclusion.

利用郧则葬皂鄄杂糟澡皂蚤凿贼过程袁在自变量集合中选择对判别分类解释性最强的信息袁删除对分类无显著解释作用的信息以及重复解释的信息袁并把挑选出来的解释变量集合变换成若干直交变量援一方面实现了判别分析模型中的变量筛选袁同时也解决了自变量多重共线条件下的有效建模问题援在选入变量的过程中运用云统计量检验变量的判别作用袁更容易被统计应用人员所接受援为了说明所提算法的合理性和有效性袁以云蚤泽澡藻则判别分析建模为例袁通过仿真数据建模取得了合理准确的分析结论援

Domaines: Statistiques [stat]

Huiwen Wang, Meiling Chen, Gilbert Saporta. Variable selection in discriminant analysis based on Gram-Schmidt process. Journal of Beijing University of Aeronautics and Astronautics, 2011, 37 (8), pp.958-961. ⟨hal-02507976⟩