Python：Kernel based PCA手工代码实现

'''
Auther: DeniuHe
Date:2021-03-16
'''
import numpy as np
from sklearn import datasets
from sklearn.metrics.pairwise import rbf_kernel
from sklearn.decomposition import KernelPCA
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt

class Kernel_PCA():
    def __init__(self,X):
        self.X = X
        self.N = self.X.shape[0]
        self.K = rbf_kernel(X=self.X,gamma=15)
        self.In = np.ones((self.N,self.N)) * (1/self.N)
        self.K_hat = self.K - self.K @ self.In - self.In @ self.K + self.In @ self.K @ self.In
        self.L, self.U = np.linalg.eig(self.K_hat)


    # dimensionality reduction
    def DR(self,n_component):
        a = [i for i in range(n_component)]
        ord_ids = np.flipud(np.argsort(self.L))
        tar_ids = ord_ids[a]
        trans_X = self.K_hat @ self.U[:,tar_ids] / np.sqrt(self.L[tar_ids])
        return trans_X



if __name__ == '__main__':
    X, y = datasets.make_moons(noise=0.1,n_samples=300,random_state=100)
    # plt.scatter(X[:,0],X[:,1],c=y)
    # plt.show()
    # X_train,X_test,y_train,y_test = train_test_split(X,y,test_size=0.1,random_state=100)
    kpca = Kernel_PCA(X=X)
    trans_X = kpca.DR(n_component=2)
    plt.scatter(trans_X[:,0],trans_X[:,1],c=y)
    plt.show()
    X, y = datasets.make_moons(noise=0.1, n_samples=300, random_state=100)
    diaobao = KernelPCA(kernel='rbf',n_components=2,gamma=15)
    X_diao = diaobao.fit_transform(X)
    plt.scatter(X_diao[:,0],X_diao[:,1],c=y)
    plt.show()

两个月亮

 

本文代码效果 （gamma=15）

sklearn调包效果（参数保持相同，即gamma=15） 

 

结果方向貌似反了，不过不要紧。 

完整版 ：

'''
Auther: DeniuHe
Date:2021-03-16
'''
import numpy as np
from sklearn import datasets
from sklearn.metrics.pairwise import rbf_kernel
from sklearn.decomposition import KernelPCA
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt

class Kernel_PCA():
    def __init__(self,X,gamma=None):
        self.X = X
        self.N = self.X.shape[0]
        self.K = rbf_kernel(X=self.X,gamma=gamma)
        self.In = np.ones((self.N,self.N)) * (1/self.N)
        self.K_hat = self.K - self.K @ self.In - self.In @ self.K + self.In @ self.K @ self.In
        self.L, self.U = np.linalg.eig(self.K_hat)


    # dimensionality reduction
    def DR(self,n_component):
        a = [i for i in range(n_component)]
        ord_ids = np.flipud(np.argsort(self.L))
        tar_ids = ord_ids[a]
        trans_X = self.K_hat @ self.U[:,tar_ids] / np.sqrt(self.L[tar_ids])
        return trans_X

    def DR_for_test_instance(self,X,n_component,gamma):
        a = [i for i in range(n_component)]
        ord_ids = np.flipud(np.argsort(self.L))
        tar_ids = ord_ids[a]
        ##
        nTest = X.shape[0]
        Inl = np.ones((self.N,nTest))*(1/self.N)
        K_test = rbf_kernel(X=self.X,Y=X,gamma=gamma)
        print("K_test::",K_test.T.shape)
        K_test_hat = K_test.T - K_test.T @ self.In - Inl.T @ self.K + Inl.T @ self.K @ self.In
        print("K_test_hat::",K_test_hat.shape)
        trans_X = K_test_hat @ self.U[:, tar_ids] / np.sqrt(self.L[tar_ids])
        print("trans_X::",trans_X.shape)
        return trans_X




if __name__ == '__main__':
    X, y = datasets.make_circles(noise=0.05,factor=0.3, n_samples=300,random_state=100)
    plt.scatter(X[:,0],X[:,1],c=y)
    plt.show()
    gamma = 9

    # kpca = Kernel_PCA(X=X,gamma=gamma)
    # trans_X = kpca.DR(n_component=2)
    # plt.scatter(trans_X[:,0],trans_X[:,1],c=y)
    # plt.show()
    #
    # diaobao = KernelPCA(kernel='rbf',n_components=2,gamma=gamma)
    # X_diao = diaobao.fit_transform(X)
    # plt.scatter(X_diao[:,0],X_diao[:,1],c=y)
    # plt.show()

    X_train,X_test,y_train,y_test = train_test_split(X,y,test_size=0.1,random_state=100)
    kpca = Kernel_PCA(X=X_train,gamma=gamma)
    trans_X = kpca.DR(n_component=2)
    plt.scatter(trans_X[:,0],trans_X[:,1],c=y_train)
    plt.show()
    trans_X_test = kpca.DR_for_test_instance(X=X_test,n_component=2,gamma=10)
    plt.scatter(trans_X_test[:,0],trans_X_test[:,1],c=y_test)
    plt.show()

 

kernel PCA 公式推导 推荐博文：

https://zhuanlan.zhihu.com/p/59775730