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训练的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维单词向量作为特征进行二进制分类。
以下是培训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例如下:
0 87
1 64
dtype: int64
所以这是一个略微不平衡的类数据集,比例为0:1 = 1:35
我现在为SVM和Random Forest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GridSearchCV for SVM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
我为RF提供了42个超参数值组合,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,如果你看一下SVM和RF的输出,我的训练精度接近99%,但测试精度和验证精度并不接近训练精度。这应该建议过度拟合,但我使用网格搜索和随机森林进行超参数调整通常也不会过度拟合。
那么可能导致这种低测试/验证准确度的原因是什么?
此外,来自ROC图的两者的AUC都非常接近0.96。所以AUC很好,准确性很差我能理解课堂失衡问题可能会起作用。但后来我在两者中使用了类权重参数来处理它。那么我的测试和验证准确度是否与培训不相上下?
这可能是因为测试数据较少(65)?
编辑:
以下是我如何进行功能的标准化。
# Standardizing the data with zero mean and Unit standard deviation of each feature (columns)
from sklearn import preprocessing
# Getting the standardizing scaler to be used for any new data too
scaler = preprocessing.StandardScaler().fit(train_X_norm)
train_X_std=scaler.transform(train_X_norm)
## Using the same transformation fitted on training data to transform the test data.
test_X_std=scaler.transform(test_X_norm)
我只在训练数据上使用标准化程序,然后使用它来转换测试数据。一个人不应该包括测试数据来计算每个特征的标准偏差和平均值,因为这将是作弊。
但即便这样做,我的测试精度也低于非标准化数据的测试精度。这很奇怪
答案 0 :(得分:1)
这不是过度拟合的问题。
您的训练集可以涵盖所有情况吗?实际上,如果您使用神经网络来适应这种分类问题,即使使用随机字嵌入,您也可以获得完美的训练结果。但是训练集和测试集(实际情况)之间没有相关性,因此测试结果将与随机分类一样糟糕。
你的情况很相似。您随机选择一些样本作为测试样本,并将遗骸留作训练集。但是,您是否可以确保测试集中的每个样本都包含训练集中的相关(类似)样本?一般来说,答案是否定的,因此测试结果通常低于训练结果。相关性越低,测试结果越低。
此外,生产结果也会低于测试结果,测试装置只是模拟生产环境。
所以不要担心你的程序,它工作正常。