机器学习预测时label错位对未来数据做预测

前言

　　这篇文章时承继上一篇机器学习经典模型使用归一化的影响。这次又有了新的任务，通过将label错位来对未来数据做预测。

实验过程

　　使用不同的归一化方法，不同得模型将测试集label错位，计算出MSE的大小；

　　不断增大错位的数据的个数，并计算出MSE，并画图。通过比较MSE（均方误差，mean-square error）的大小来得出结论

过程及结果

数据处理(和上一篇的处理方式相同)：

1 test_sort_data = sort_data[:5000]
2 test_sort_target = sort_target[:5000]
3 
4 sort_data1 = _sort_data[5000:16060]
5 sort_data2 = _sort_data[16060:]
6 sort_target1 = _sort_target[5000:16060]
7 sort_target2 = _sort_target[16060:]
View Code

完整数据处理代码：

 1 #按时间排序
 2 sort_data = data.sort_values(by = 'time',ascending = True)
 3 
 4 sort_data.reset_index(inplace = True,drop = True)
 5 target = data['T1AOMW_AV']
 6 sort_target = sort_data['T1AOMW_AV']
 7 del data['T1AOMW_AV']
 8 del sort_data['T1AOMW_AV']
 9 
10 from sklearn.model_selection import train_test_split
11 test_sort_data = sort_data[16160:]
12 test_sort_target = sort_target[16160:]
13 
14 _sort_data = sort_data[:16160]
15 _sort_target = sort_target[:16160]
16 
17 from sklearn.model_selection import train_test_split
18 test_sort_data = sort_data[:5000]
19 test_sort_target = sort_target[:5000]
20 
21 sort_data1 = _sort_data[5000:16060]
22 sort_data2 = _sort_data[16060:]
23 sort_target1 = _sort_target[5000:16060]
24 sort_target2 = _sort_target[16060:]
25 
26 import scipy.stats as stats
27 dict_corr = {
28     'spearman' : [],
29     'pearson' : [],
30     'kendall' : [],
31     'columns' : []
32 }
33 
34 for i in data.columns:
35     corr_pear,pval = stats.pearsonr(sort_data[i],sort_target)
36     corr_spear,pval = stats.spearmanr(sort_data[i],sort_target)
37     corr_kendall,pval = stats.kendalltau(sort_data[i],sort_target)
38     
39     dict_corr['pearson'].append(abs(corr_pear))
40     dict_corr['spearman'].append(abs(corr_spear))
41     dict_corr['kendall'].append(abs(corr_kendall))
42     
43     dict_corr['columns'].append(i)
44     
45 # 筛选新属性  
46 dict_corr =pd.DataFrame(dict_corr)
47 dict_corr.describe()
View Code

选取25%以上的；

1 new_fea = list(dict_corr[(dict_corr['pearson']>0.41) & (dict_corr['spearman']>0.45) & (dict_corr['kendall']>0.29)]['columns'].values)
View Code

包含下面的用来画图：

1 import matplotlib.pyplot as plt 
2 lr_plt=[]
3 ridge_plt=[]
4 svr_plt=[]
5 RF_plt=[]
View Code

正常的计算mse（label没有移动）：

 1 from sklearn.linear_model import LinearRegression,Lasso,Ridge
 2 from sklearn.preprocessing import MinMaxScaler,StandardScaler,MaxAbsScaler
 3 from sklearn.metrics import mean_squared_error as mse
 4 from sklearn.svm import SVR
 5 from sklearn.ensemble import RandomForestRegressor
 6 import xgboost as xgb
 7 #最大最小归一化
 8 mm = MinMaxScaler()
 9 
10 lr = Lasso(alpha=0.5)
11 lr.fit(mm.fit_transform(sort_data1[new_fea]), sort_target1)
12 lr_ans = lr.predict(mm.transform(sort_data2[new_fea]))
13 lr_mse=mse(lr_ans,sort_target2)
14 lr_plt.append(lr_mse)
15 print('lr:',lr_mse)
16 
17 ridge = Ridge(alpha=0.5)
18 ridge.fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
19 ridge_ans = ridge.predict(mm.transform(sort_data2[new_fea]))
20 ridge_mse=mse(ridge_ans,sort_target2)
21 ridge_plt.append(ridge_mse)
22 print('ridge:',ridge_mse)
23 
24 svr = SVR(kernel='rbf',C=100,epsilon=0.1).fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
25 svr_ans = svr.predict(mm.transform(sort_data2[new_fea]))
26 svr_mse=mse(svr_ans,sort_target2)
27 svr_plt.append(svr_mse)
28 print('svr:',svr_mse)
29 
30 estimator_RF = RandomForestRegressor().fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
31 predict_RF = estimator_RF.predict(mm.transform(sort_data2[new_fea]))
32 RF_mse=mse(predict_RF,sort_target2)
33 RF_plt.append(RF_mse)
34 print('RF:',RF_mse)
35 
36 bst = xgb.XGBRegressor(learning_rate=0.1, n_estimators=550, max_depth=4, min_child_weight=5, seed=0,
37                              subsample=0.7, colsample_bytree=0.7, gamma=0.1, reg_alpha=1, reg_lambda=1)
38 bst.fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
39 bst_ans = bst.predict(mm.transform(sort_data2[new_fea]))
40 print('bst:',mse(bst_ans,sort_target2))
View Code

先让label移动5个：

1 change_sort_data2 = sort_data2.shift(periods=5,axis=0)
2 change_sort_target2 = sort_target2.shift(periods=-5,axis=0)
3 change_sort_data2.dropna(inplace=True)
4 change_sort_target2.dropna(inplace=True)
View Code

让label以5的倍数移动：

 1 mm = MinMaxScaler()
 2 
 3 for i in range(0,45,5):
 4     print(i)
 5     lr = Lasso(alpha=0.5)
 6     lr.fit(mm.fit_transform(sort_data1[new_fea]), sort_target1)
 7     lr_ans = lr.predict(mm.transform(change_sort_data2[new_fea]))
 8     lr_mse=mse(lr_ans,change_sort_target2)
 9     lr_plt.append(lr_mse)
10     print('lr:',lr_mse)
11     
12     ridge = Ridge(alpha=0.5)
13     ridge.fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
14     ridge_ans = ridge.predict(mm.transform(change_sort_data2[new_fea]))
15     ridge_mse=mse(ridge_ans,change_sort_target2)
16     ridge_plt.append(ridge_mse)
17     print('ridge:',ridge_mse)
18     
19     svr = SVR(kernel='rbf',C=100,epsilon=0.1).fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
20     svr_ans = svr.predict(mm.transform(change_sort_data2[new_fea]))
21     svr_mse=mse(svr_ans,change_sort_target2)
22     svr_plt.append(svr_mse)
23     print('svr:',svr_mse)
24     
25     estimator_RF = RandomForestRegressor().fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
26     predict_RF = estimator_RF.predict(mm.transform(change_sort_data2[new_fea]))
27     RF_mse=mse(predict_RF,change_sort_target2)
28     RF_plt.append(RF_mse)
29     print('RF:',RF_mse)
30     
31 #     bst = xgb.XGBRegressor(learning_rate=0.1, n_estimators=550, max_depth=4, min_child_weight=5, seed=0,
32 #                              subsample=0.7, colsample_bytree=0.7, gamma=0.1, reg_alpha=1, reg_lambda=1)
33 #     bst.fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
34 #     bst_ans = bst.predict(mm.transform(change_sort_data2[new_fea]))
35 #     print('bst:',mse(bst_ans,change_sort_target2))
36     
37     change_sort_target2=change_sort_target2.shift(periods=-5,axis=0)
38     change_sort_target2.dropna(inplace=True)
39     change_sort_data2 = change_sort_data2.shift(periods=5,axis=0)
40     change_sort_data2.dropna(inplace=True)
View Code

结果如图：

然后就是画图了；

1 plt.plot(x,lr_plt,label='lr',color='r',marker='o')
2 plt.plot(x,ridge_plt,label='ridge',color='b',marker='o')
3 plt.plot(x,svr_plt,label='svr',color='g',marker='o')
4 plt.plot(x,RF_plt,label='RF',color='y',marker='o')
5 plt.legend()
6 plt.show()
View Code

舍去lr，并扩大纵坐标：

1 #plt.plot(x,lr_plt,label='lr',color='r',marker='o')
2 plt.plot(x,ridge_plt,label='ridge',color='b',marker='o')
3 plt.plot(x,svr_plt,label='svr',color='g',marker='o')
4 plt.plot(x,RF_plt,label='RF',color='y',marker='o')
5 plt.legend()
6 plt.show()
View Code

其他模型只需将MinMaxScaler改为MaxAbsScaler,standarScaler即可；

总的来说，label的移动会使得mse增加，大约在label=10时候差异最小，结果最理想；

posted @ 2018-11-23 02:09  StarHai  阅读(567)  评论(0)  编辑  收藏  举报