主成分分析法PCA（二）：python实现

三：算法代码

这里我将代码分享给大家，但大家使用时请附上来源！！

import numpy as np;
class PCA:
    def __init__(self,rate=0.85,is_normal=False):
        self.rate = rate;
        self.is_normal=is_normal;
    def average(self,x,ready=False):
        aver = lambda x:sum(x)/len(x);
        x_1 = x if ready else np.array(x).T;
        x_aver = [aver(a) for a in x_1];
        x_aver = np.array(x_aver);
        #相当于x_aver = np.mean(x_1,axis=1);
        return x_aver;

    def std(self,x,ready=False):
        x_1 = x if ready else np.array(x).T;
        s=[];
        for a in x_1:
            s.append(np.std(a));
        s=np.array(s);
        return s;

    def cov(self,x,ready=False):
        x_1 = x if ready else np.array(x).T;
        return np.cov(x_1);

    def eigen(self,cov_mat):
        '''vals: 特征值，list类型
        vecs：特征向量，按行摆放，一行即一个向量,ndarray类型'''
        vals,vecs = np.linalg.eig(cov_mat); # 输出类型为ndarray类型
        vecs = vecs.T;
        vals_new=sorted(vals,reverse=True); # 按从大到小排列
        vecs_new=np.array([vecs[vals.tolist().index(a)] for a in vals_new]);
        return vals_new,vecs_new;

    def contribution(self,vals):
        sum_val = sum(vals);
        ctr = [a/sum_val for a in vals];
        return ctr;

    def pca_k(self,ctr,rate):
        sum_ctr = 0;
        for i in range(len(ctr)):
            sum_ctr += ctr[i];
            if sum_ctr>=rate:
                return i+1;

    def normalization(self,X,ready=False):
        X = X if ready else np.array(X);
        X_new = (X-self.aver)/self.s;
        return X_new;

    def fit(self,X):
        x = np.array(X).T;
        if self.is_normal:
            self.aver = self.average(x,True); # 求平均
            self.s = self.std(x,True);        # 求标准差
            x = self.normalization(x.T,True).T; # 进行标准化处理
        self.cov_mat = self.cov(x,True);        # 求协方差矩阵
        self.vals,self.vecs = self.eigen(self.cov_mat); # 求特征值和特征向量
        self.ctr = self.contribution(self.vals); # 求主成分的方差贡献率
        self.k = self.pca_k(self.ctr,self.rate); # 或取主成分的k值
        self.covv = self.vecs[:self.k]; # 组成关系矩阵
        if self.is_normal:
            lmk = np.sqrt(np.array(self.vals)).T;
            self.covv = self.covv*lmk;

    def transfer(self,X):
        x_1 = np.array(X);
        if self.is_normal:
            x_1 = self.normalization(x_1,True);
        y=[];
        for x in x_1:
            y.append(np.matmul(self.covv,x).T);
        y=np.array(y).tolist();
        return y;

调用上述算法：

x1=[149.5,162.5,162.7,162.2,156.5,156.1,172.0,173.2,159.5,157.7];
x2=[69.5,77.0,78.5,87.5,74.5,74.5,76.5,81.5,74.5,79.0];
x3=[38.5,55.5,50.8,65.5,49.0,45.5,51.0,59.5,43.5,53.5];
x=[[a,x2[x1.index(a)],x3[x1.index(a)]] for a in x1];

pca=PCA(rate=0.9,is_normal=False);
pca.fit(x);
print(pca.cov_mat,'-------------------------',sep='\n');
print(pca.vals,pca.vecs,'-------------------------',sep='\n');
print(pca.k,pca.covv,'-------------------------',sep='\n');
print(np.array(pca.transfer(x[:2])));

结果：

[[51.74544444 18.98666667 34.41922222]
 [18.98666667 23.45555556 36.19555556]
 [34.41922222 36.19555556 61.69566667]]
-------------------------
[110.00413886434858, 25.32447972824003, 1.5680480740780194]
[[ 0.55915657  0.42128705  0.71404562]
 [ 0.82767368 -0.33348264 -0.45138188]
 [-0.04796048 -0.84338992  0.53515721]]
-------------------------
2
[[ 0.55915657  0.42128705  0.71404562]
 [ 0.82767368 -0.33348264 -0.45138188]]
-------------------------
[[140.36411374  83.18196923]
 [162.93157761  83.76711524]]