线性回归原理及两种代码实现

原理

前提假设

给定一组由输入x和输出y构成的数据集D = {(x₁,y₁),(x₂,y₂),(x₃，y₃),…（x_m，x_m)},其中x_i=(x_i1;x_i2;x_i3…；x_id),y_i∈R。

原理阐述

线性回归是通过训练学习得到一个线性模型y来最大限度地根据输入x拟合输出y。即$\hat{y}=w{x}_i+b$ 确定参数w和b使得拟合值$\hat{y}$和真实值y的均方误差尽可能小。

代码

基于numpy的代码实现

import numpy as np
def linear_loss(X,y,w,b):
    ##回归模型主体
    ##输入 ：X变量矩阵 y标签向量 w参数权重矩阵 b偏置
    ##输出 dw db loss y_hat
    num_train = X.shape[0]         ##得到数据量
    num_feature = X.shape[1]       ##得到特征数
    y_hat = np.dot(X,w)+b          ##线性模型y^=wx+b
    loss = np.sum((y_hat-y)**2)/num_train    ##计算均方损失
    dw = np.dot(X.T,(y_hat-y))/num_train     ##计算梯度
    db = np.sum((y_hat-y))/num_train         ##计算梯度
    return y_hat,loss,dw,db


def initialize_params(dims):                     
    ##初始化函数
    ## 输入dims：数据维度 数据特征数量
    ##输出： 初始权重矩阵w   初始偏置参数b
    w = np.zeros((dims,1))
    b = 0
    return w,b    


def linear_train(X,y,learning_rate=0.01,epochs=20000):
    ##模型最优化过程
    ##输入 数据矩阵X 标签值y 学习率 迭代次数
    ##输出 每次迭代后的均方损失loss_every  优化后的参数字典params   迭代后的梯度
    loss_every = []     ##存储均方损失
    w,b = initialize_params(X.shape[1])    ##得到初始参数
    for i in range(1,epochs):              ##梯度下降法
        y_hat,loss,dw,db = linear_loss(X,y,w,b)
        w += -learning_rate * dw
        b += -learning_rate * db
        loss_every.append(loss)
        if i%20000 == 0:        ##每训练20000次，打印一次迭代次数i和损失loss
            print("epoch %d loss %f"%(i,loss))   
        params = {'w':w,                              ##将w和b存到字典里
                  'b':b}
        grads = {'dw':dw,
                 'db':db}          
    return loss_every,params,grads


##调用数据集 并分为训练集和测试集
from sklearn.datasets import load_diabetes
from sklearn.utils import shuffle
diabetes = load_diabetes()
data,target = diabetes.data,diabetes.target
X,y = shuffle(data,target,random_state=13)   ##打乱数据集
offset = int(X.shape[0]*0.8)        ##获得数据数量*0.8  再取整
X_train,y_train = X[:offset],y[:offset]
X_test,y_test = X[offset:],y[offset:]
y_train = y_train.reshape((-1,1))
y_test = y_test.reshape((-1,1))
print("X_train's shape:",X_train.shape)
print("X_test's shape:",X_test.shape)

##模型训练
loss_every,params,grads = linear_train(X_train,y_train,0.01,200000)
print(params)


##模型预测
def predict(X,params):
    ## 输入 数据矩阵X   含w，b参数字典params
    ## 输出 y的预测值
    w = params['w']
    b = params['b'] ##字典值调用
    y_pred = np.dot(X,w)+b
    return y_pred
    y_pred = predict(X_test,params)


def r2_score(y_test,y_pred):
    ##模型效果评估
    ## 输入y的测试集的值和y的预测值
    ## 输出r2
    y_avg = np.mean(y_test)
    ss_tot = np.sum((y_test-y_avg)**2)
    ss_res = np.sum((y_test-y_pred)**2)
    r2 = 1-(ss_res/ss_tot)
    return r2
print(r2_score(y_test,y_pred))
        

基于sklearn的代码实现

from sklearn import linear_model
from sklearn.metrics import mean_squared_error,r2_score
from sklearn.datasets import load_diabetes
from sklearn.utils import shuffle

diabetes = load_diabetes()   ##加载数据集
data,target = diabetes.data,diabetes.target
X,y = shuffle(data,target,random_state=13)   ##打乱数据集
offset = int(X.shape[0]*0.8)        ##获得数据量*0.8 再取整
X_train,y_train = X[:offset],y[:offset]
X_test,y_test = X[offset:],y[offset:]
y_train = y_train.reshape((-1,1))
y_test = y_test.reshape((-1,1))

model = linear_model.LinearRegression()     ##定义模型
model.fit(X_train,y_train)                  ##拟合模型
predict = model.predict(X_test)             ##模型预测
print("均方误差：%.2f"% mean_squared_error(y_test,predict))  ##模型评估
print("R²：%.2f"% r2_score(y_test,predict))