【深度之眼吴恩达机器学习第四期】笔记（十一）

最新推荐文章于 2021-11-23 18:10:16 发布

树天先森

最新推荐文章于 2021-11-23 18:10:16 发布

阅读量339

点赞数

分类专栏：吴恩达机器学习文章标签：机器学习推荐系统

本文链接：https://blog.csdn.net/qq_40923177/article/details/105155373

版权

吴恩达机器学习专栏收录该内容

12 篇文章 3 订阅

订阅专栏

编程

导包并展示数据

import numpy as np
import matplotlib.pyplot as plt
from scipy.io import loadmat

data = loadmat('data/ex8_movies.mat')
data

# 用户对电影的评分矩阵
Y = data['Y']
# 用户对某部电影是否评分的矩阵，已评分为1，未评分为0
R = data['R']
# ((1682, 943), (1682, 943))
Y.shape, R.shape

# 对第一部电影评分的平均值为3.2061068702290076
Y[1,np.where(R[1,:]==1)[0]].mean()

fig, ax = plt.subplots(figsize=(12,12))
ax.imshow(Y)
ax.set_xlabel('Users')
ax.set_ylabel('Movies')
# 自动调整子图参数（坐标轴标签、刻度标签以及标题的部分），使之填充整个图像区域。
fig.tight_layout()
plt.show()

Y代表用户对电影的评分（从1到5，0代表没有进行评分）。
R代表用户有没有对某部电影评分，0代表没有评分，1代表有评分。
在这里插入图片描述
可以看到Y的最后几行几乎都是黑色的，代表这些电影几乎没有用户对其进行评分。

计算损失和梯度

def cost(params, Y, R, num_features, l):
# INPUT：参数params（也就是X,theta），数据与标签Y,R，特征个数num_features，正则化参数lambda
# OUTPUT：梯度grad，代价函数J

    # 初始化变量
    Y = np.matrix(Y)  # (1682, 943)
    R = np.matrix(R)  # (1682, 943)
    num_movies = Y.shape[0]
    num_users = Y.shape[1]
    J = 0
    X_grad = np.zeros(X.shape)
    Theta_grad = np.zeros(Theta.shape)    
 
    # 把参数params还原回矩阵的形式
    X = np.matrix(np.reshape(params[:num_movies * num_features], (num_movies, num_features)))
    Theta = np.matrix(np.reshape(params[num_movies * num_features:], (num_users, num_features)))
    
    # 计算代价函数
    error = np.multiply((X * Theta.T) - Y, R)
    squared_error = np.power(error, 2)
    J = (1. / 2) * np.sum(squared_error)    
    # 加入正则项
    J = J + l*np.sum(np.power(Theta,2))/2
    J = J + l*np.sum(np.power(X,2))/2
        
    # 计算包含正则项的梯度
    X_grad = error * Theta+l*X
    Theta_grad = error.T * X+l*Theta    
    # 把矩阵拉伸成向量
    grad = np.concatenate((np.ravel(X_grad), np.ravel(Theta_grad)))    
    return J, grad

使用少量数据来检查函数是否正确

params_data = loadmat('data/ex8_movieParams.mat')
X = params_data['X']
Theta = params_data['Theta']
# ((1682, 10), (943, 10))
X.shape, Theta.shape

users = 4
movies = 5
features = 3

# 获取数据中的一部分
X_sub = X[:movies, :features]
Theta_sub = Theta[:users, :features]
Y_sub = Y[:movies, :users]
R_sub = R[:movies, :users]
params = np.concatenate((np.ravel(X_sub), np.ravel(Theta_sub)))

J, grad = cost(params, Y_sub, R_sub, features, 1.5)
J, grad
'''(31.34405624427422,
 array([ -0.95596339,   6.97535514,  -0.10861109,   0.60308088,
          2.77421145,   0.25839822,   0.12985616,   4.0898522 ,
         -0.89247334,   0.29684395,   1.06300933,   0.66738144,
          0.60252677,   4.90185327,  -0.19747928, -10.13985478,
          2.10136256,  -6.76563628,  -2.29347024,   0.48244098,
         -2.99791422,  -0.64787484,  -0.71820673,   1.27006666,
          1.09289758,  -0.40784086,   0.49026541]))'''

加载电影标题的数据

movie_idx = {}
f = open('data/movie_ids.txt',encoding= 'gbk')
for line in f:
    # line:1 Toy Story (1995)
    # tokens:['1', 'Toy', 'Story', '(1995)\n']
    tokens = line.split(' ')
    # tokens:['1', 'Toy', 'Story', '(1995)']
    tokens[-1] = tokens[-1][:-1]
    # int(tokens[0]) - 1 = 0
    # ' '.join(tokens[1:])='Toy Story (1995)'
    movie_idx[int(tokens[0]) - 1] = ' '.join(tokens[1:])

# 'Toy Story (1995)'
movie_idx[0]

这是某个新用户对电影的评分

ratings = np.zeros((1682, 1))

ratings[0] = 4
ratings[6] = 3
ratings[11] = 5
ratings[53] = 4
ratings[63] = 5
ratings[65] = 3
ratings[68] = 5
ratings[97] = 2
ratings[182] = 4
ratings[225] = 5
ratings[354] = 5

print('Rated {0} with {1} stars.'.format(movie_idx[0], str(int(ratings[0]))))
print('Rated {0} with {1} stars.'.format(movie_idx[6], str(int(ratings[6]))))
print('Rated {0} with {1} stars.'.format(movie_idx[11], str(int(ratings[11]))))
print('Rated {0} with {1} stars.'.format(movie_idx[53], str(int(ratings[53]))))
print('Rated {0} with {1} stars.'.format(movie_idx[63], str(int(ratings[63]))))
print('Rated {0} with {1} stars.'.format(movie_idx[65], str(int(ratings[65]))))
print('Rated {0} with {1} stars.'.format(movie_idx[68], str(int(ratings[68]))))
print('Rated {0} with {1} stars.'.format(movie_idx[97], str(int(ratings[97]))))
print('Rated {0} with {1} stars.'.format(movie_idx[182], str(int(ratings[182]))))
print('Rated {0} with {1} stars.'.format(movie_idx[225], str(int(ratings[225]))))
print('Rated {0} with {1} stars.'.format(movie_idx[354], str(int(ratings[354]))))
'''Rated Toy Story (1995) with 4 stars.
Rated Twelve Monkeys (1995) with 3 stars.
Rated Usual Suspects, The (1995) with 5 stars.
Rated Outbreak (1995) with 4 stars.
Rated Shawshank Redemption, The (1994) with 5 stars.
Rated While You Were Sleeping (1995) with 3 stars.
Rated Forrest Gump (1994) with 5 stars.
Rated Silence of the Lambs, The (1991) with 2 stars.
Rated Alien (1979) with 4 stars.
Rated Die Hard 2 (1990) with 5 stars.
Rated Sphere (1998) with 5 stars.'''

把新用户的数据添加到已有的数据中，原来有943个用户，现在有944个用户。进行均值归一化。

R = data['R']
Y = data['Y']
Y = np.append(Y, ratings, axis=1)
R = np.append(R, ratings != 0, axis=1)
# ((1682, 944), (1682, 944), (1682, 1))
Y.shape, R.shape, ratings.shape

movies = Y.shape[0]  # 1682
users = Y.shape[1]  # 944
features = 10
learning_rate = 10.

# 随机初始化X和theta
X = np.random.random(size=(movies, features))
Theta = np.random.random(size=(users, features))
params = np.concatenate((np.ravel(X), np.ravel(Theta)))
# ((1682, 10), (944, 10), (26260,))
X.shape, Theta.shape, params.shape

# 均值归一化
Ymean = np.zeros((movies, 1))
Ynorm = np.zeros((movies, users))
for i in range(movies):
    # 对于第i部电影，获取评分的下标
    idx = np.where(R[i,:] == 1)[0]
    # 计算该电影的平均分
    Ymean[i] = Y[i,idx].mean()
    # 均值归一化
    Ynorm[i,idx] = Y[i,idx] - Ymean[i]

# 因为减去均值后，均值约等于0，Ynorm.mean()得到5.5070364565159845e-19是对的
Ynorm.mean()

使用均值归一化后的Y训练模型

from scipy.optimize import minimize
fmin = minimize(fun=cost, x0=params, args=(Ynorm, R, features, learning_rate), 
                method='CG', jac=True, options={'maxiter': 100})
fmin
'''
     fun: 38959.17920754673
     jac: array([ 0.06997272,  0.06133648,  0.07338806, ...,  0.00534678, -0.00109083, -0.00755571])
 message: 'Maximum number of iterations has been exceeded.'
    nfev: 152
     nit: 100
    njev: 152
  status: 1
 success: False
       x: array([ 0.73899214,  0.49007446,  0.33694708, ...,  0.05745608, -0.13162036,  0.06442625])
'''

预测分数并给用户推荐电影

# 把向量还原回矩阵
X = np.matrix(np.reshape(fmin.x[:movies * features], (movies, features)))
Theta = np.matrix(np.reshape(fmin.x[movies * features:], (users, features)))
# ((1682, 10), (944, 10))
X.shape, Theta.shape

# 预测电影的评分
predictions = X * Theta.T
# 只计算新加用户的预测
my_preds = predictions[:, -1] + Ymean
# (1682, 1)
my_preds.shape
# 把已经看过的电影排除
mark = ratings==0
my_preds = np.multiply(my_preds,mark)

# np.sort()从上到下是从小到大的，[::-1]表示倒序
sorted_preds = np.sort(my_preds, axis=0)[::-1]
# 找到预测评分最高的10部电影
sorted_preds[:10]
'''
matrix([[5.00000099],
        [4.99999988],
        [4.99999952],
        [4.99999922],
        [4.99999884],
        [4.99999857],
        [4.99999849],
        [4.99999834],
        [4.99999759],
        [4.99999755]])
'''

idx = np.argsort(my_preds, axis=0)[::-1]
idx
'''
Out[47]:matrix([[1188],
        [1598],
        [1121],
        ...,
        [  53],
        [ 354],
        [   0]], dtype=int64)
'''

print("Top 10 movie predictions:")
for i in range(10):
    j = int(idx[i])
    print('Predicted rating of {0} for movie {1}.'.format(str(float(my_preds[j])), movie_idx[j]))
'''
Top 10 movie predictions:
Predicted rating of 5.000000993596591 for movie Prefontaine (1997).
Predicted rating of 4.999999875214552 for movie Someone Else's America (1995).
Predicted rating of 4.99999952030098 for movie They Made Me a Criminal (1939).
Predicted rating of 4.999999216853268 for movie Aiqing wansui (1994).
Predicted rating of 4.999998837107815 for movie Saint of Fort Washington, The (1993).
Predicted rating of 4.999998569336983 for movie Great Day in Harlem, A (1994).
Predicted rating of 4.999998494568234 for movie Star Kid (1997).
Predicted rating of 4.999998340436829 for movie Santa with Muscles (1996).
Predicted rating of 4.999997590707319 for movie Entertaining Angels: The Dorothy Day Story (1996).
Predicted rating of 4.9999975535076215 for movie Marlene Dietrich: Shadow and Light (1996) .
'''