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)：

 test_sort_data = sort_data[:]
 test_sort_target = sort_target[:]

 sort_data1 = _sort_data[:]
 sort_data2 = _sort_data[:]
 sort_target1 = _sort_target[:]
 sort_target2 = _sort_target[:]

完整数据处理代码：

 #按时间排序
 sort_data = data.sort_values(by = 'time',ascending = True)

 sort_data.reset_index(inplace = True,drop = True)
 target = data['T1AOMW_AV']
 sort_target = sort_data['T1AOMW_AV']
 del data['T1AOMW_AV']
 del sort_data['T1AOMW_AV']

 from sklearn.model_selection import train_test_split
 test_sort_data = sort_data[:]
 test_sort_target = sort_target[:]

 _sort_data = sort_data[:]
 _sort_target = sort_target[:]

 from sklearn.model_selection import train_test_split
 test_sort_data = sort_data[:]
 test_sort_target = sort_target[:]

 sort_data1 = _sort_data[:]
 sort_data2 = _sort_data[:]
 sort_target1 = _sort_target[:]
 sort_target2 = _sort_target[:]

 import scipy.stats as stats
 dict_corr = {
     'spearman' : [],
     'pearson' : [],
     'kendall' : [],
     'columns' : []
 }

 for i in data.columns:
     corr_pear,pval = stats.pearsonr(sort_data[i],sort_target)
     corr_spear,pval = stats.spearmanr(sort_data[i],sort_target)
     corr_kendall,pval = stats.kendalltau(sort_data[i],sort_target)

     dict_corr['pearson'].append(abs(corr_pear))
     dict_corr['spearman'].append(abs(corr_spear))
     dict_corr['kendall'].append(abs(corr_kendall))

     dict_corr['columns'].append(i)

 # 筛选新属性
 dict_corr =pd.DataFrame(dict_corr)
 dict_corr.describe()

选取25%以上的；

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

包含下面的用来画图：

 import matplotlib.pyplot as plt
 lr_plt=[]
 ridge_plt=[]
 svr_plt=[]
 RF_plt=[]

正常的计算mse（label没有移动）：

 from sklearn.linear_model import LinearRegression,Lasso,Ridge
 from sklearn.preprocessing import MinMaxScaler,StandardScaler,MaxAbsScaler
 from sklearn.metrics import mean_squared_error as mse
 from sklearn.svm import SVR
 from sklearn.ensemble import RandomForestRegressor
 import xgboost as xgb
 #最大最小归一化
 mm = MinMaxScaler()

 lr = Lasso(alpha=0.5)
 lr.fit(mm.fit_transform(sort_data1[new_fea]), sort_target1)
 lr_ans = lr.predict(mm.transform(sort_data2[new_fea]))
 lr_mse=mse(lr_ans,sort_target2)
 lr_plt.append(lr_mse)
 print('lr:',lr_mse)

 ridge = Ridge(alpha=0.5)
 ridge.fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
 ridge_ans = ridge.predict(mm.transform(sort_data2[new_fea]))
 ridge_mse=mse(ridge_ans,sort_target2)
 ridge_plt.append(ridge_mse)
 print('ridge:',ridge_mse)

 svr = SVR(kernel='rbf',C=,epsilon=0.1).fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
 svr_ans = svr.predict(mm.transform(sort_data2[new_fea]))
 svr_mse=mse(svr_ans,sort_target2)
 svr_plt.append(svr_mse)
 print('svr:',svr_mse)

 estimator_RF = RandomForestRegressor().fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
 predict_RF = estimator_RF.predict(mm.transform(sort_data2[new_fea]))
 RF_mse=mse(predict_RF,sort_target2)
 RF_plt.append(RF_mse)
 print('RF:',RF_mse)

 bst = xgb.XGBRegressor(learning_rate=0.1, n_estimators=, max_depth=, min_child_weight=, seed=,
                              subsample=0.7, colsample_bytree=0.7, gamma=0.1, reg_alpha=, reg_lambda=)
 bst.fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
 bst_ans = bst.predict(mm.transform(sort_data2[new_fea]))
 print('bst:',mse(bst_ans,sort_target2))

先让label移动5个：

 change_sort_data2 = sort_data2.shift(periods=,axis=)
 change_sort_target2 = sort_target2.shift(periods=-,axis=)
 change_sort_data2.dropna(inplace=True)
 change_sort_target2.dropna(inplace=True)

让label以5的倍数移动：

 mm = MinMaxScaler()

 for i in range(,,):
     print(i)
     lr = Lasso(alpha=0.5)
     lr.fit(mm.fit_transform(sort_data1[new_fea]), sort_target1)
     lr_ans = lr.predict(mm.transform(change_sort_data2[new_fea]))
     lr_mse=mse(lr_ans,change_sort_target2)
     lr_plt.append(lr_mse)
     print('lr:',lr_mse)

     ridge = Ridge(alpha=0.5)
     ridge.fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
     ridge_ans = ridge.predict(mm.transform(change_sort_data2[new_fea]))
     ridge_mse=mse(ridge_ans,change_sort_target2)
     ridge_plt.append(ridge_mse)
     print('ridge:',ridge_mse)

     svr = SVR(kernel='rbf',C=,epsilon=0.1).fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
     svr_ans = svr.predict(mm.transform(change_sort_data2[new_fea]))
     svr_mse=mse(svr_ans,change_sort_target2)
     svr_plt.append(svr_mse)
     print('svr:',svr_mse)

     estimator_RF = RandomForestRegressor().fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
     predict_RF = estimator_RF.predict(mm.transform(change_sort_data2[new_fea]))
     RF_mse=mse(predict_RF,change_sort_target2)
     RF_plt.append(RF_mse)
     print('RF:',RF_mse)

 #     bst = xgb.XGBRegressor(learning_rate=0.1, n_estimators=, max_depth=, min_child_weight=, seed=,
 #                              subsample=0.7, colsample_bytree=0.7, gamma=0.1, reg_alpha=, reg_lambda=)
 #     bst.fit(mm.fit_transform(sort_data1[new_fea]),sort_target1)
 #     bst_ans = bst.predict(mm.transform(change_sort_data2[new_fea]))
 #     print('bst:',mse(bst_ans,change_sort_target2))

     change_sort_target2=change_sort_target2.shift(periods=-,axis=)
     change_sort_target2.dropna(inplace=True)
     change_sort_data2 = change_sort_data2.shift(periods=,axis=)
     change_sort_data2.dropna(inplace=True)

结果如图：

然后就是画图了；

 plt.plot(x,lr_plt,label='lr',color='r',marker='o')
 plt.plot(x,ridge_plt,label='ridge',color='b',marker='o')
 plt.plot(x,svr_plt,label='svr',color='g',marker='o')
 plt.plot(x,RF_plt,label='RF',color='y',marker='o')
 plt.legend()
 plt.show()

舍去lr，并扩大纵坐标：

 #plt.plot(x,lr_plt,label='lr',color='r',marker='o')
 plt.plot(x,ridge_plt,label='ridge',color='b',marker='o')
 plt.plot(x,svr_plt,label='svr',color='g',marker='o')
 plt.plot(x,RF_plt,label='RF',color='y',marker='o')
 plt.legend()
 plt.show()

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；