Re:Automatic Generation of Regular Expressions from Examples
by haffi112 on Monday October 22, 2012 @07:42AM
(#4682679)
Then you might have the problem of over-fitting the training set.
See here for example: http://en.wikipedia.org/wiki/Overfitting
Re:Automatic Generation of Regular Expressions from Examples
by msellout on Monday October 22, 2012 @10:01AM
(#4683135)
More training data helps prevent overfitting.
Re:Automatic Generation of Regular Expressions from Examples
by StavrosK on Monday October 22, 2012 @08:28AM
(#4682818)
You use the validation set precisely to avoid overfitting. I agree that a 90/10 split would be more reasonable.
Re:Automatic Generation of Regular Expressions from Examples
by ZoFreX on Monday October 22, 2012 @10:04AM
(#4683155)
You could also use 10-fold cross validation[1] or similar where you rotate the training/validation sets.
[1]: http://en.wikipedia.org/wiki/Cross-validation_%28statistics%...
Re:Automatic Generation of Regular Expressions from Examples
by icelancer on Monday October 22, 2012 @06:59PM
(#4685986)
Also natively supported in WEKA. (Not that WEKA is easy to use.)
Re:Automatic Generation of Regular Expressions from Examples
by StavrosK on Monday October 22, 2012 @10:09AM
(#4683186)
Yep, that's another good way.
Re:Automatic Generation of Regular Expressions from Examples
by Cyranix on Monday October 22, 2012 @11:40AM
(#4683618)
The other replies, directly discussing overfitting, are spot on. However, I did want to point out explicitly that the stated assumption of representative sample data can be a fundamental challenge, not to be taken lightly.
Why divide the data into sets like that? Assuming the sample data is representative, wouldn't it be better to use much more for training, say 90% with 10% for validation?