I was given a situation to predict the validity of the logistic regression model when it was found that certain values of a heavily weighted feature were found to be erroneously multiplied by 1000. For example, the feature has values 20.000, 78.000, 56.000, 10.000 and so on. However, few values were found to be missing the decimal point and were therefore scaled up by 1000. Will the results of the previously fit model be valid in such a situation and why ? Also, how will the model be affected by the wrongly scaled up values.
1 Answer
$\begingroup$
$\endgroup$
2
In my opinion, the model will still be valid because even if the range of the predictor variable is high in the case of Logistic Regression, the algorithm itself will improvise itself by reducing the value of the "coefficient" corresponding to this variable.
You can read more about this here and here
Dhanyawad.
-
$\begingroup$ I understand that the algorithm will itself adjust the feature coefficient, however, in this case, the scale of the variable in consideration is erroneously boosted, the kind we might not see in the test set, will the model still be valid ? $\endgroup$ Commented Jan 28, 2021 at 21:30
-
$\begingroup$ @AnkitaTalwar Let suppose your training dataset feature in consideration was supposed to have a range of [1, 100] but erroneously some of them missed a decimal value. So, the range has become [1, 60000]. Since the model has already calibrated itself with respect to the range it will still be "valid". To visualize the change please refer to this colab notebook link $\endgroup$– MrityuCommented Jan 29, 2021 at 7:22