线性回归是最典型的回归问题,其目标值与所有的特征之间存在线性关系。线性回归于逻辑回归类似,不同的是,逻辑回归在线性回归的基础上加了逻辑函数,从而将线性回归的值从实数域映射到了0-1,通过设定阀值,便实现了回归的0-1分类,即二分类。
线性回归函数$Y=XW$,其中Y是1*n维向量,X是n*m矩阵,W是m*1的系数矩阵。线性回归采用平方损失函数,至于为什么采用平方损失函数,是因为平方损失函数采用的是最小二乘法的思想,其残差满足正态分布的最大似然估计,详情可百度。
线性回归损失函数:${{l}_{w}}=\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-X_{i}W \right)}^{2}}}$,其中$X_i$是1*m维向量,W是m*1维向量。
线性回归的解法有很多种,如直接最小二乘法,梯度下降法和牛顿法等。
1. 最小二乘法
直接最小二乘法是利用矩阵变换,直接求得系数向量W的矩阵解,过程如下。
预测函数为:$Y=XW$,其损失函数表示为:${{Y-XW}^{T}}(Y-XW)$
对W求导可得:$\frac{d}{dW}{{Y-XW}^{T}}(Y-XW)={{X}^{T}}(Y-XW)$,其求导过程需要矩阵求导的知识。
另导数为0,得到:$W={{\left( {{X}^{T}}X \right)}^{-1}}{{X}^{T}}Y$
2. 梯度下降法
线性回归损失函数:${{l}_{w}}=\sum\limits_{i=1}^{n}{{{\left( {{y}_{i}}-X_{i}W \right)}^{2}}}$,要求损失函数的最小值,对参数W求偏导,得:
\[\frac{\partial {{\text{l}}_{w}}}{\partial W}=\sum\limits_{i=1}^{n}{\left( {{y}_{i}}-{{X}_{i}}W \right)}\cdot {{X}_{i}}\]
由上式可知,参数W的梯度和逻辑回归中类似,是每个样本的残差值乘以样本对应的值,然后累加起来,第$j$个参数的梯度是对所有样本第$j$特征执行上述操作。
线性回归最小二乘和梯度下降法Python代码如下:
# -*- coding: utf-8 -*- """ Created on Fri Jan 19 13:29:14 2018 @author: zhang """ import numpy as np from sklearn.datasets import load_boston import matplotlib.pyplot as plt from sklearn.cross_validation import train_test_split from sklearn import preprocessing """ 多元线性回归需要对各变量进行标准化,因为在求系数wj梯度时,每个样本计算值与其标签值的差要与每个样本对应的第j个属性值相乘,然后求和 因此,如果属性值之间的差异太大,会造成系数无法收敛 """ # 最小二乘法直接求解权重系数 def least_square(train_x, train_y): """ input:训练数据(样本*属性)和标签 """ weights = (train_x.T * train_x).I * train_x.T * train_y return weights # 梯度下降算法 def gradient_descent(train_x, train_y, maxCycle, alpha): numSamples, numFeatures = np.shape(train_x) weights = np.zeros((numFeatures,1)) for i in range(maxCycle): h = train_x * weights err = h - train_y weights = weights - (alpha * err.T * train_x).T return weights def stochastic_gradient_descent(train_x, train_y, maxCycle, alpha): numSamples, numFeatures = np.shape(train_x) weights = np.zeros((numFeatures,1)) for i in range(maxCycle): for j in range(numSamples): h = train_x[j,:] * weights err = h - train_y[j,0] weights = weights - (alpha * err.T * train_x[j,:]).T return weights def load_data(): boston = load_boston() data = boston.data label = boston.target return data, label def show_results(predict_y, test_y): plt.scatter(np.array(test_y), np.array(predict_y), marker=‘x‘, s= 30, c=‘red‘) # 画图的数据需要是数组而不能是矩阵 plt.plot(np.arange(0,50),np.arange(0,50)) plt.xlabel("original_label") plt.ylabel("predict_label") plt.title("LinerRegression") plt.show() if __name__ == "__main__": data, label = load_data() data = preprocessing.normalize(data.T).T train_x, test_x, train_y, test_y = train_test_split(data, label, train_size = 0.75, random_state = 33) train_x = np.mat(train_x) test_x = np.mat(test_x) train_y = np.mat(train_y).T # (3,)转为矩阵变为行向量了,需要转置 test_y = np.mat(test_y).T # weights = least_square(train_x, train_y) # predict_y = test_x * weights # show_results(predict_y, test_y) # weights = gradient_descent(train_x, train_y, 1000, 0.01) predict_y = test_x * weights show_results(predict_y, test_y) # weights = stochastic_gradient_descent(train_x, train_y, 100, 0.01) # predict_y = test_x * weights # show_results(predict_y, test_y)