Baktaawar
Reputation: 7490
Possible reason for Lower Test Accuracy but high AUC score
Lets say I have a dataset like below:
word label_numeric
0 active 0
1 adventurous 0
2 aggressive 0
3 aggressively 0
4 ambitious 0
I use a word2Vec trained model and convert each word into their word vector of 300 dimensions. This is how it looks now.
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]:
I have two labels 0 and 1. I am now doing a Binary classification with 300 dimension word vectors as features.
Here is the details of training and testing count:
# 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
Now my label ratio in training data is as below:
0 87
1 64
dtype: int64
So it's slightly imbalanced class dataset with ratio of 0:1=1:35
I now do a GridSearchCV for both SVM and Random Forest. In both the algo, i put
class_weights={1:1.35,0:1}
to take into account the class imbalance problem in machine learning.
Here is my GridSearchCV function:
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)
I get the following as result for SVM.
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
The parameter grid for hyperparamter values passed to GridSearchCV for SVM is:
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]}]
I ran Random Forest also:
Here is the result:
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
I had 42 combination of hyper parameter values for RF as below:
grid_rf={'n_estimators': [30,100,250,500,600,900], 'max_depth':[2,4,7,8,9,10,13]}
Now if you look at both the outputs for SVM and RF, my training accuracy is like close to 99% but test accuracy and validation accuracy is not close to training accuracy. This should suggest Overfitting, but I did the hyper parameter tuning using Grid Search and Random Forest generally doesn't overfit too.
So what could be causing this low test/validation accuracy?
Also the AUC of both from ROC plot is very good close to 0.96. So AUC is good, and accuracy is bad I can understand class imbalance issue might be in play. But then I took care of that using class weights parameter in both. So then also my test and validation accuracy is not comparable to training?
Can this be because of less test data (65)?
Edit:
Here is how I did the standardization of features.
# 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)
I fit the standardizer on the training data only and then use that to transform the test data. One shouldnt be including the test data to calculate the standard deviation and mean of each feature as that would be cheating.
But even after doing this, my test accuracy falls below what I had for non standardized data. That is strange
Upvotes: 0
Views: 2192
Answers (1)
Sraw
Reputation: 20214
This is not a problem about overfitting.
Can your training set cover all situation? Actually, if you are using neural network to fit this classification problem, you can also get a perfect training result even though using random word embedding. But there is no relevance between training set and test set(real situation), so the test result will be as bad as random classification.
Your situation is similar. You randomly choose some samples as test sample, and leave the remains as training set. But could you ensure that every sample in your test set has relevant(similar) samples in your training set? Generally, the answer is NO, so the test result is usually lower than training result. The lower relevance, the lower test result will be.
Further, result in production will be also lower than test result, test set is just a simulation of production environment.
So don't worry about your program, it works fine.
Upvotes: 1
Related Questions
- Why roc_auc produces weird results in sklearn?
- SVM performance not consistent with AUC score
- Finding AUC score for SVM model
- Why am I getting an almost perfect test accuracy?
- Getting a low ROC AUC score but a high accuracy
- sklearn svm area under ROC less than 0.5 for training data
- 100% Accuracy using SVC classification, something must be wrong?
- why model is giving high accuracy of 84% but very low AUC 13%?
- Accuracy of model is 0.86 while AUC is 0.50?
- scikit-learn roc_auc_score() returns accuracy values