假设我有一个如下数据集：
word label_numeric
0 active 0
1 adventurous 0
2 aggressive 0
3 aggressively 0
4 ambitious 0
我使用 word2Vec 训练模型并将每个单词转换为 300 维的单词向量。这就是它现在的样子。
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 label
0 0.058594 -0.016235 -0.174805 0.072266 -0.201172 0.073242 -0.074219 -0.149414 0.245117 -0.050049 -0.016357 -0.147461 -0.003311 0.071289 -0.008545 -0.179688 0.001686 -0.009949 -0.036621 0.048096 -0.033447 0.105957 -0.490234 0.249023 -0.199219 -0.025635 -0.248047 0.136719 -0.068848 -0.320312 0.259766 -0.053223 0.154297 -0.050537 0.110840 0.027100 0.000412 -0.133789 0.077148 0.058838 0.230469 -0.033203 -0.179688 -0.125977 -0.166992 -0.110352 -0.365234 -0.330078 -0.021729 -0.076660 0.124023 -0.107910 -0.051758 0.127930 0.192383 0.025024 0.033691 -0.386719 -0.006195 -0.074219 -0.175781 -0.088379 -0.341797 0.145508 -0.051758 0.099609 0.020874 -0.042969 -0.145508 0.090332 0.096191 0.061768 0.209961 0.314453 -0.080078 -0.304688 0.238281 -0.060791 0.146484 0.041504 -0.113281 0.019409 0.328125 0.300781 -0.153320 -0.174805 -0.347656 -0.002167 0.115723 0.104004 0.012817 -0.175781 0.088867 -0.291016 -0.092773 0.144531 -0.006256 -0.066406 -0.145508 -0.182617 -0.144531 0.074707 -0.157227 -0.025513 -0.013977 -0.289062 0.051514 -0.010559 0.121582 0.072754 0.005188 -0.162109 -0.246094 0.002014 -0.072266 -0.026733 0.143555 0.067383 0.398438 -0.212891 0.029663 -0.041748 -0.005157 0.337891 -0.192383 -0.135742 0.226562 -0.033691 -0.188477 0.322266 0.136719 -0.058594 -0.068359 0.136719 0.029175 -0.152344 -0.086426 0.021729 -0.005524 0.115723 0.106445 0.257812 0.000546 -0.161133 -0.046875 -0.049805 -0.058594 -0.110840 0.029907 -0.322266 -0.032715 -0.136719 -0.148438 0.125977 -0.205078 0.027222 -0.005219 -0.188477 0.318359 0.002792 0.155273 0.261719 -0.043457 0.113281 0.142578 0.170898 -0.202148 0.028687 0.239258 0.033203 -0.330078 -0.003647 -0.054199 -0.142578 0.201172 0.053467 -0.249023 -0.180664 0.147461 -0.036865 -0.015259 -0.107910 -0.134766 0.052002 0.109863 0.067871 0.022705 0.058838 -0.189453 -0.093262 -0.043945 -0.009216 0.020386 -0.232422 -0.083008 0.062500 0.016479 0.033936 0.041016 0.049805 0.071289 0.076660 -0.003937 -0.261719 -0.198242 -0.269531 -0.035889 -0.249023 -0.023071 -0.091797 -0.093750 0.192383 -0.376953 0.170898 0.027832 0.023438 0.047363 -0.051270 0.020386 -0.029663 0.128906 0.044434 -0.199219 0.060547 0.138672 0.104980 0.314453 -0.125000 -0.075684 0.088379 0.109863 -0.058594 0.063477 -0.120117 -0.177734 0.017700 0.112793 -0.161133 -0.188477 -0.102051 -0.068848 -0.073730 0.168945 -0.042236 -0.024536 0.128906 -0.066406 -0.020996 0.087891 -0.224609 0.025146 -0.054932 -0.102539 -0.020142 0.123047 -0.171875 0.195312 -0.203125 -0.265625 -0.026367 0.154297 -0.235352 0.092773 0.032715 0.177734 0.063477 -0.168945 0.153320 -0.182617 0.101074 0.074219 0.031250 -0.038086 0.037598 0.035400 -0.150391 -0.108398 -0.071289 -0.080078 0.078613 0.022705 0.148438 -0.098633 -0.032471 0.083984 0.031494 -0.052002 -0.062988 0.316406 -0.105957 0.026733 0.018921 0.026855 -0.176758 -0.088379 0.127930 -0.104980 0.206055 -0.003296 0.184570 0
1 -0.068359 0.076660 -0.224609 0.292969 0.054688 -0.069824 0.028809 0.090332 -0.160156 0.080566 0.289062 -0.005615 0.074219 -0.071289 0.069824 0.032715 -0.036133 0.043457 0.084961 0.224609 -0.001160 0.100098 -0.090820 0.209961 0.101074 0.009949 0.038818 0.151367 0.209961 -0.157227 0.118652 0.247070 0.090332 0.244141 0.125000 -0.253906 0.204102 -0.234375 0.118652 -0.000603 0.253906 -0.146484 -0.077148 0.180664 -0.110840 0.018677 -0.113770 0.159180 0.245117 -0.033447 -0.041748 0.246094 0.018677 0.034180 0.103516 0.087891 0.339844 -0.357422 -0.230469 -0.051758 -0.038574 -0.281250 -0.218750 -0.210938 -0.150391 -0.040283 -0.049072 -0.292969 0.151367 0.143555 0.048340 -0.194336 -0.027344 0.038574 -0.086426 -0.003036 -0.095215 0.062500 -0.098145 0.085938 -0.099609 0.046875 0.039551 0.182617 -0.142578 0.189453 -0.261719 0.030273 0.056152 0.123535 -0.082520 -0.075684 -0.267578 0.014832 0.047852 -0.012451 0.131836 0.240234 -0.107910 -0.316406 0.081055 0.092285 0.014771 0.211914 0.062500 -0.143555 0.412109 -0.210938 -0.064453 -0.193359 0.051025 0.027954 0.026367 -0.109375 0.020752 -0.124512 0.198242 -0.105469 0.250000 -0.071289 -0.065430 -0.139648 -0.032959 0.386719 -0.185547 -0.166992 0.036621 0.001389 -0.090820 0.030396 -0.249023 -0.047363 -0.013245 0.318359 -0.150391 0.048340 -0.037354 0.125000 -0.053711 0.562500 0.005463 -0.067383 -0.345703 0.214844 0.044678 0.170898 -0.218750 0.243164 -0.165039 -0.259766 -0.158203 -0.275391 -0.138672 0.080566 -0.212891 -0.238281 -0.075684 0.015320 0.089844 -0.052490 0.031738 0.339844 0.035400 0.212891 0.127930 -0.033447 0.234375 0.130859 -0.209961 -0.106445 -0.236328 0.047607 -0.153320 -0.075195 0.048340 0.133789 -0.085449 0.122070 -0.187500 -0.172852 -0.137695 -0.392578 -0.028809 -0.177734 -0.131836 -0.141602 0.071777 -0.118652 -0.072754 -0.081543 -0.070312 0.033447 0.124023 -0.088379 -0.130859 0.131836 -0.010437 0.247070 -0.287109 0.077637 0.033203 0.032959 -0.136719 -0.079590 0.051758 -0.045898 -0.131836 -0.326172 -0.202148 -0.033203 -0.176758 0.180664 -0.148438 0.227539 -0.212891 -0.143555 0.273438 0.134766 -0.261719 0.073242 -0.054688 0.027466 0.126953 0.234375 0.097168 0.259766 0.253906 -0.170898 -0.189453 0.239258 -0.173828 0.024536 0.002090 0.101074 0.351562 0.174805 0.162109 -0.146484 -0.103516 -0.037354 0.065430 -0.104004 0.108398 0.296875 0.172852 0.078613 -0.209961 -0.133789 0.037354 -0.125977 0.172852 -0.102539 0.034424 0.095215 0.158203 -0.291016 -0.047852 -0.161133 -0.024414 -0.162109 -0.161133 0.109375 0.003372 0.218750 -0.022339 0.057861 -0.351562 -0.113770 -0.247070 -0.108398 0.097656 0.083008 0.357422 0.347656 0.341797 -0.031006 0.056885 0.114746 0.083008 0.192383 0.335938 0.154297 -0.244141 -0.445312 0.166992 0.396484 -0.132812 0.077148 -0.108398 0.131836 0.063477 0.001221 -0.219727 -0.062988 -0.137695 -0.133789 0.223633 -0.069336 0.163086 0.236328 0
2 -0.003067 0.219727 -0.082520 0.255859 -0.209961 -0.117188 0.109863 0.107422 0.059570 0.007233 0.059082 -0.152344 0.208984 -0.095703 -0.096680 -0.312500 -0.154297 0.024780 0.032471 0.250000 0.090820 0.017944 0.105957 0.133789 -0.122070 0.199219 -0.073730 -0.142578 0.203125 0.047607 0.222656 0.019531 0.026123 -0.138672 0.061768 0.120605 -0.008789 -0.047852 0.269531 -0.182617 0.566406 -0.218750 -0.043457 -0.051270 -0.273438 -0.084961 -0.240234 -0.158203 0.221680 -0.043457 0.308594 0.221680 -0.112305 -0.014343 0.070312 0.174805 -0.090332 -0.384766 0.003281 -0.002808 -0.273438 -0.116211 -0.542969 -0.008057 -0.137695 0.209961 0.231445 -0.008484 -0.092285 0.226562 -0.021851 -0.083984 0.069336 0.277344 -0.217773 0.057129 0.269531 0.218750 0.137695 0.093750 -0.101562 0.281250 0.029785 0.126953 0.066406 -0.019775 -0.287109 0.267578 0.195312 -0.135742 0.012207 0.048828 -0.237305 0.101562 0.206055 -0.091309 -0.085938 0.112305 -0.008423 -0.037109 0.099121 0.018433 -0.108398 0.031982 0.202148 -0.273438 -0.007874 -0.179688 0.025879 -0.046387 -0.172852 -0.202148 -0.086426 -0.028564 -0.033447 -0.047852 0.184570 -0.146484 0.109863 -0.243164 -0.251953 -0.000456 -0.073730 0.199219 -0.248047 -0.265625 0.261719 0.003693 0.092285 -0.111816 -0.118652 -0.320312 0.121582 0.127930 -0.127930 -0.087402 0.229492 0.040527 -0.121094 0.233398 0.052734 0.213867 -0.111328 -0.030884 -0.084961 0.054932 -0.068848 0.133789 -0.121582 -0.235352 -0.031982 0.062500 -0.137695 0.244141 -0.070312 -0.090820 -0.050781 0.041748 0.166992 0.200195 0.016724 0.292969 0.023682 -0.232422 -0.113281 -0.032959 0.038330 -0.357422 0.187500 -0.034180 -0.157227 -0.213867 0.007233 0.136719 0.018433 0.040771 0.089355 0.162109 -0.051514 -0.109863 -0.142578 -0.292969 -0.043945 0.200195 -0.079102 -0.007172 0.131836 0.206055 -0.125977 -0.092285 0.118652 -0.042236 -0.054443 -0.082520 -0.238281 -0.078125 0.052979 0.003601 -0.045166 0.126953 0.064453 0.296875 0.145508 -0.006378 0.015869 -0.070312 0.036377 -0.277344 0.038574 -0.112793 -0.224609 0.171875 -0.184570 0.062500 0.142578 -0.170898 0.189453 -0.067871 -0.239258 -0.110840 -0.043213 0.089844 0.069824 0.012512 0.162109 -0.194336 0.419922 -0.116699 0.170898 0.119141 -0.189453 0.102051 0.055420 0.026245 0.008545 0.052246 -0.088379 -0.236328 -0.041016 -0.125000 -0.051514 0.020020 0.051758 -0.137695 0.206055 -0.029297 -0.106445 -0.039062 0.285156 -0.018677 0.265625 -0.072266 -0.090820 -0.030640 -0.112793 -0.181641 -0.000690 -0.171875 -0.115234 -0.179688 0.114746 0.032227 -0.016235 -0.063477 0.054688 -0.033691 -0.242188 -0.292969 -0.229492 0.067871 0.006378 0.345703 0.024780 0.148438 0.119629 0.121582 0.024780 0.086914 0.066895 0.181641 0.120605 0.234375 0.034180 -0.306641 -0.124512 0.145508 0.025269 -0.138672 0.353516 -0.227539 -0.082520 -0.035645 0.066895 -0.085938 -0.159180 -0.087402 0.186523 0.289062 -0.075195 0.050781 0
In [223]:
我有两个标签 0 和 1。我现在正在使用 300 维词向量作为特征进行二进制分类。
以下是训练和测试计数的详细信息：
# Splitting the dataset to train test
from sklearn.cross_validation import train_test_split
train_X, test_X,train_Y,test_Y = train_test_split(jpsa_X_norm,jpsa_Y, test_size=0.30, random_state=42)
print("Total Sample size in Training {}\n".format(train_X.shape[0]))
print("Total Sample size in Test {}".format(test_X.shape[0]))
Total Sample size in Training 151
Total Sample size in Test 65
现在我在训练数据中的标签比率如下：
0 87
1 64
dtype: int64
所以它是稍微不平衡的类数据集，比例为 0:1=1:35
我现在为 SVM 和随机森林做一个 GridSearchCV。在这两个算法中，我把
class_weights={1:1.35,0:1}
考虑机器学习中的类不平衡问题。
这是我的 GridSearchCV 函数：
def grid_search(self):
"""This function does Cross Validation using Grid Search
"""
from sklearn.model_selection import GridSearchCV
self.g_cv = GridSearchCV(estimator=self.estimator,param_grid=self.param_grid,cv=5)
self.g_cv.fit(self.train_X,self.train_Y)
我得到以下 SVM 的结果。
The mean train scores are [ 0.57615906 0.57615906 0.57615906 0.57615906 0.93874475 0.57615906
0.57615906 0.57615906 1. 0.94867633 0.57615906 0.57615906
1. 1. 0.950343 0.57615906 0.81777921 0.99668044
1. 1. ]
The mean validation scores are [ 0.57615894 0.57615894 0.57615894 0.57615894 0.87417219 0.57615894
0.57615894 0.57615894 0.8807947 0.8807947 0.57615894 0.57615894
0.86754967 0.87417219 0.88741722 0.57615894 0.70860927 0.90728477
0.87417219 0.87417219]
The score on held out data is: 0.9072847682119205
Parameters for Best Score : {'C': 1, 'kernel': 'linear'}
The accuracy of svm on test data is: 0.8769230769230769
Classification Metrics for svm :
precision recall f1-score support
0 0.87 0.92 0.89 37
1 0.88 0.82 0.85 28
avg / total 0.88 0.88 0.88 65
传递给 SVM 的 GridSearchCV 的超参数值的参数网格是：
grid_svm=[{'kernel': ['rbf'], 'gamma': [1e-1,1e-2,1e-3,1e-4],\
'C': [0.1, 1, 10, 100]},\
{'kernel': ['linear'], 'C': [0.1,1,10,100]}]
我也跑了随机森林：
结果如下：
The mean train scores are [ 0.99009597 1. 0.99833333 1. 0.99833333 1.
0.99834711 1. 1. 1. 1. 1. 1.
1. 1. 1. 1. 1. 1. 1.
1. 1. 1. 1. 1. 1. 1.
1. 1. 1. 1. 1. 1. 1.
1. 1. 1. 1. 1. 1. 1.
1. ]
The mean validation scores are [ 0.79470199 0.85430464 0.8807947 0.87417219 0.8807947 0.85430464
0.83443709 0.82781457 0.86754967 0.84768212 0.88741722 0.87417219
0.81456954 0.86092715 0.85430464 0.83443709 0.8410596 0.8410596
0.83443709 0.86092715 0.85430464 0.83443709 0.84768212 0.82781457
0.82781457 0.82119205 0.85430464 0.81456954 0.82781457 0.85430464
0.82781457 0.84768212 0.83443709 0.86092715 0.87417219 0.86754967
0.86092715 0.86092715 0.8410596 0.86754967 0.86754967 0.8410596 ]
The score on held out data is: 0.8874172185430463
Parameters for Best Score : {'max_depth': 4, 'n_estimators': 600}
The accuracy of rf on test data is: 0.8307692307692308
Classification Metrics for rf :
precision recall f1-score support
0 0.77 1.00 0.87 37
1 1.00 0.61 0.76 28
avg / total 0.87 0.83 0.82 65
我有 42 个 RF 的超参数值组合，如下所示：
grid_rf={'n_estimators': [30,100,250,500,600,900], 'max_depth':[2,4,7,8,9,10,13]}
现在，如果您同时查看 SVM 和 RF 的输出，我的训练准确度接近 99%，但测试准确度和验证准确度并不接近训练准确度。这应该表明过度拟合，但我使用网格搜索和随机森林进行了超参数调整，通常也不会过度拟合。
那么是什么导致了这种低测试/验证准确性呢？
ROC 图的 AUC 也非常好，接近 0.96。所以 AUC 很好，但准确度很差，我可以理解类不平衡问题可能在起作用。但是后来我在两者中都使用了类权重参数来解决这个问题。那么我的测试和验证准确性也无法与培训相提并论吗？
我还添加了更多数据，所以现在我有 2000 个 0 和 1000 个 1。我在每个算法中使用 scikit learn class_weight 选项中的“balanced”选项来实现类不平衡
但是现在如果我获取更多数据，这是随机森林对更多数据的结果。
The mean train scores are [ 0.70347493 0.73347328 0.74070792 0.74368715 0.74609988 0.74772955
0.7476584 0.78035322 0.80624038 0.81432687 0.8194324 0.81581485
0.81773002 0.81929078 0.9497877 0.96858105 0.97283788 0.97524883
0.9759579 0.97567365 0.9751775 0.97851051 0.99099354 0.99248265
0.99489341 0.99468108 0.99538994 0.99595762 0.98999975 0.99794336
0.99872325 0.99893632 0.99872348 0.99914909 0.99907804 0.99687948
0.99957447 0.99978721 0.99957452 0.99978728 0.99971639 0.99978728
0.99985806 1. 1. 1. 1. 1. 1. ]
The mean validation scores are [ 0.68765957 0.71460993 0.7222695 0.71829787 0.71744681 0.72453901
0.71971631 0.7248227 0.73191489 0.74439716 0.74638298 0.74524823
0.74695035 0.74297872 0.75716312 0.77730496 0.78468085 0.78382979
0.79120567 0.78609929 0.7906383 0.75120567 0.77531915 0.78808511
0.78780142 0.79035461 0.79234043 0.78808511 0.75716312 0.7693617
0.78297872 0.78553191 0.78609929 0.77957447 0.78269504 0.75234043
0.77673759 0.77021277 0.7764539 0.76879433 0.77134752 0.77673759
0.74241135 0.75148936 0.75375887 0.75375887 0.75432624 0.75829787
0.75205674]
The score on held out data is: 0.7923404255319149
Hyper-Parameters for Best Score : {'max_depth': 8, 'n_estimators': 700}
The accuracy of rf on test data is: 0.8022486772486772
Classification Metrics for rf :
precision recall f1-score support
0 0.83 0.90 0.86 956
1 0.71 0.64 0.67 433
2 0.92 0.62 0.74 123
avg / total 0.80 0.80 0.80 1512
这似乎将准确率从 82% 降低到 80%。为什么会这么冷？如果数据在增加，那么为什么更多的数据准确性会下降？结果表明训练精度为 1，但验证和测试接近 t0 0.8。这是为什么？是否存在过度拟合，因为验证错误很高并且训练错误很低，但是随机森林通常不会过度拟合。
或者这可能是因为添加的新数据可能是噪音而不是真正的标签？所以它降低了旧小数据的准确性度量？