3
$\begingroup$

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?

enter image description here

I also added more data so now I have 2000 0's and 1000 1's. I use the option "balanced" in scikit learn class_weight option in each algorithm for class imbalance

But now if I take more data, here is the result of random forest on more data.

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

This seem to have decreased the accuracy from 82 to 80%. Why cold that be? If data is increasing then why for more data accuracy goes down? The result show that training accuracy is 1 but validation and test is close t0 0.8. Why is that? Is there something overfitting since validation error is high and train error is low but then Random Forest generally don't overfit that well.

Or Can this be, because the new data added might be noise and not true labels? So it reduces the accuracy measure on old small data?.

$\endgroup$
3
  • $\begingroup$ Please adjust your original question to focus on what you really need. For instance in your current question you ask "This seem to have decreased the accuracy from 82 to 80%. Why cold that be?" - I have tried to answer you why that is not a problem, and you say that actually you are not worried about this. To get an answer to the thing you think is important, you need to make it clear what is the important issue in your question. Don't mention your tests with small number of samples at all, they hide your question, get rid of that part (use the edit link). $\endgroup$ May 10 '17 at 19:43
  • $\begingroup$ When I put that question I had two parts to it: 1) On small data sample is there a concern of the difference in training and test accuracy. Which as I suspected too, you answered there is no issue as test sample is very low. 2) My second part is what about when the sample size is more as I indicated in the end of the question. That part is the answer I am seeking now. So question has two parts hence both the data size mention $\endgroup$
    – Baktaawar
    May 10 '17 at 19:52
  • $\begingroup$ I think it is probably too much for a single question. Too many parts. It is ok to ask more than one question, and to take more time over things. However, in this case I don't see much use to keep the first part of the question. I still suggest to edit the question to make it shorter and more focused on a single issue. $\endgroup$ May 10 '17 at 20:16
1
$\begingroup$

I'm not sure if this is correct since I'm still new but I read (and I'm not very good at reading) that measuring accuracy on classification models isn't ideal. It's better to measure precision and recall which is what you are already doing. Have you tried using the confusion matrix? This will tell you how many times the model classified something as 0 or 1. For classification, I just use sklearn's classification report. It'll give you the Precision, Recall, and F1 scores.

$\endgroup$
0
$\begingroup$

Your initial sample size is relatively small for machine learning. That is going to result in high variance estimates, different versions of the datasets might have different results on the evaluation metric.

As far as evaluation metric performance going down as you increase sample size, that again is a high variance with small sample size issue. Your initial small sample happens to more amendable to modeling (probably due to randomness). As you increase sample size, it lowers the variance and in this case also lowers the bias.

Be cautious about resampling and resplitting your data repeatedly and looking for the best solution. It can be tempting to mistakenly drawn conclusions about the data, the best model, and machine learning in general.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.