- reference
PCA:
Examples (proved) PCA functioning https://www.jianshu.com/p/0227aa77425f
Data pretreatment before PCA: to the center, standardization (theoretical win) https://blog.csdn.net/u010182633/article/details/45918737
Practice on the training set and test set (details on the operation) http://wenda.chinahadoop.cn/question/5926
Others are summarized https://blog.csdn.net/viewcode/article/details/8789524
LDA:
LDA principle (binary and multi-classification): https://www.cnblogs.com/jerrylead/archive/2011/04/21/2024384.html
Graphic Codes rich https://blog.csdn.net/ruthywei/article/details/83045288
- to sum up
PCA:
Effective Gaussian distribution, because the Gaussian distribution is not related to independence
Before doing needs to be decentralized and standardized (dimensionless), because the decentralized conversion will only come variance associated with the size of orthogonal transformation, behind the derivation easy, non-dimensional view of the dimension will affect the comparison .
Wherein the PCA is orthogonal
PCA selected features reserved only for dispersion of all the data useful, may not be useful for data classification, LDA is for data classification
PCA characteristic value and the covariance matrix is that all samples (to the center, after normalization) and the center (i.e., the coordinate origin) and the square of the distance
If the classification, appropriate data can be classified using the variance , such as
LDA: For classification K, K-1 down to up above dimensional space, and not necessarily orthogonal features elected
Suitable using mean data can be distinguished