0
$\begingroup$

I took a sample of my training data and balanced it and then trained my model. The results obtained are more accurate than using the whole set of train data (balanced or imbalanced). My question is: what could explain this result? (the entire set of data does not contain any noise).

Thanks in advance.

$\endgroup$
  • $\begingroup$ Outliers could be an explanation. $\endgroup$ – Simon Larsson Jun 4 at 12:11
  • $\begingroup$ "(the entire set of data does not contain any noise)" Do you mean no outliers are contained or really no statistical variation occurs? $\endgroup$ – Alex2006 Jun 4 at 13:17
  • $\begingroup$ @Alex2006, yes i am talking about outliers $\endgroup$ – Born New Jun 4 at 13:24
  • $\begingroup$ In data without outliers, you may have a problem of overfitting. The whole data always contain more noise than a subsample which leads to more precise estimates normally countered by the division through the square root of the sample size when constructing error bars of confidence intervals. Does this occurs with any subsample or just a specific one? $\endgroup$ – Alex2006 Jun 4 at 13:30
  • $\begingroup$ It occurs with a specific one $\endgroup$ – Born New Jun 4 at 13:47
0
$\begingroup$

In absence of noise and if the difference of accuracy you observe is significant, the only reason I see is that, by luck, the distribution of the training data subset happens to be closer to the distribution of your validation data than the distribution of the whole training data.
This should not happen with truly representative validation data (i.e. with training and validation data built by random split of a given dataset).

$\endgroup$

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.