Adressing overfitting:

• 减少特征
• 模型选择,自动选择变量
  但是特征信息的舍弃会导致信息的丢失

regularization:

• 保留所有特征,但是减少参数theta的值
• 在很多特征时有良好的效果

cost function

对参数惩罚,保证参数较小,防止过拟合
1. fitting well
2. theta is small

这里的lambda参数设置过大会underfitting

正则化回归

正则化回归中的只惩罚非常数项所以,将梯度下降分开:

Normal equation

正则化通过在对角加上一个数值,可是解决不可逆的问题.

逻辑回归正则化

无正则化的逻辑回归的cost function

正则化的cost

梯度下降的式子与线性的相同,不同的是h(theta)函数不同

其损失函数为:

整个迭代过程为:

__author__ = 'Chen'

from numpy import *

#calculate the cost
def @H_404_64@costFunction(X,Y,theta):
    mse = (theta * X.T - Y.T)
    return mse *mse.T
#linearReresion
def @H_404_64@linearRegresion(x,y,type=True,alpha=0.01,lambdas=0.01):

    xrow = shape(x)[0]
    xcol = shape(x)[1]
    x = matrix(x)
    Y = matrix(y)
    # fill ones
    xone = ones((xrow,1))
    X = hstack((xone,x))
    X = matrix(X)
    # normal equiation
    if type == True:
        #add regularization
        for i in range(1,xrow):
            X[i,i] += lambdas * 1
        theta = (X.T*X).I*X.T*Y
        return theta
    else:

        # gradiant
        theta = matrix(random.random(xcol+1))
        # iterations
        for iteration in range(1,10000):
                # return the cost
                print costFunction(X,theta)
                sums = 0
                #gradient method
                # adding a regularzation need to add theta(i-1)
                temptheta = theta
                temptheta[0,0] = 0

                for i in range(1,xrow):
                    sums += (theta*X[i,:].T-Y[i,:])*X[i,:]
                theta -= alpha*sums/xrow + lambdas * temptheta/xrow
        return theta



x= [[0,1,0],[0,0,1],[1,1]]
y= [[1],[2],[3],[4]]

# calculate linearRegression by normal equation
theta1 = linearRegresion(x,y)
print theta1

#gradient descent
theta2 = linearRegresion(x,False)

print theta2