When the feature selection is completed, the direct model can be trained, but may be due to the feature matrix is too large, leading to long calculation capacity, training a matter of time, thus reducing the characteristic dimension of the matrix is also essential. Common dimensionality reduction based model items other than L1 penalty, in addition to principal component analysis (PCA) and Linear Discriminant Analysis (LDA), a linear discriminant analysis classification model itself addition to the above mentioned. PCA and LDA have a lot of similarities, its essence is the original sample you want to map to a lower dimensional sample space, but PCA and LDA mapping target is not the same: PCA is to allow the sample after the map has the greatest divergence ; while the LDA is to make the sample after the mapping has the best classification performance . So PCA is an unsupervised dimensionality reduction method, and LDA dimensionality reduction method is a kind of supervision.