Linear discriminant analysis Lda parameters description
A, the parameter
1, the Solver - o hyperplane method of characteristic matrix
Singular value decomposition (Svd method
The default solver for SVD, it can perform classification and it does not depend on the covariance matrix calculation, the number of cases the function is an advantage, SVD can have the effect of compression data set, the important information to extract data set for dimension reduction processing, features very long is recommended to use this method, the downside of this approach is not to carry on the shrinkage of regularization, likely to cause excessive fitting the distortion
(2) Lsqr least-square method
Least square solution of optimal decision rules based on the discriminant function directly, it can only be used for classification of contraction (optional), because they do not perform the eigenvector estimate, therefore, the dimension does not support using transformation for cuts, can undertake shrinkage
(3) characteristics of Eigen decomposition method
Different from SVD, eigenvalue decomposition method is suitable for the N order phalanx, eigenvalue calculation Rayleigh equation solver coefficient of the optimal solution (essentially between the dispersion and the ratio of dispersion within class scatter), the solver supports both classification and dimension reduction contraction (optional), can prevent excessive regularization processing through the shrinkage fitting
Because this project dimension, only do not classify phenomenon existed fitting process, so can't choose the least square method, also compared with the feature decomposition method in the process of computing need calculate the covariance matrix calculation is bigger, SVD less amount of calculation and therefore parameter Solver selection defaults SVD
2, the Shrinkage - regularization parameter
Regularization parameters, which can enhance the generalization ability of the LDA classification, if only to dimension reduction, then you can ignore this parameter, the default value is None, neither to regularization, can choose to "auto", let the algorithm decide whether regularization, of course, also can choose different values between the [0, 1] cross validation to participate, because the project will only for dimension, so no input, that is, choose the default value of None
3, Priors - category weights
Weight category, was in a classification model can specify different categories of weight,, in turn, affects the establishment of the classification model, dimension reduction generally do not need to pay attention to this parameter, so don't input using the default
4, n_components - for the Lda dimension reduction when dropped to the dimensions of the
When LDA dimension reduction to dimensions, need to input the parameters in dimension reduction, optional values only for [1, number of categories 1) range between integer, this project will be 2 d dimension reduction to the one-dimensional data set, so here choose 1.
5, Store_covariance - storage covariance
If you want to another class of covariance matrix (the default is False), the parameter selected bool, use only in "SVD solver, our project is to klearn dataset WeiHua dimension data sets, don't need to calculate the covariance matrix, thus using the default value of False
6, Tol - floating
The default value is 0.0001, the SVD solver is used to rank estimation of threshold value,