How to use Adversarial Validation to Help Fix Overfitting

[@lakeguy65616]

at approx 9:53, you display a chart of "feature importances"... How do you determine feature importance? Thank you, GREAT VIDEO!

[@artem_isakow]

Thank you!

[@jonathanlarkin1112]

Thanks for this. Very nice video. I’m wondering about the use of AV in time series problems were you have lag and moving average features. Any special care needed in that situation?

[@arunavamaulik19]

Thank you very much for the video, very helpful. Subscribed!

[@nicholasliu-sontag1585]

Would the 'ideal' roc_auc score be 0.5?

[@welcomeaioverlords]

Yes!

[@wei_zou]

I kept wondering whether allowing people to use feature matrix in the testing data (without label) will inflate the model performance on the same testing data. Yes, people can win a kaggle contest like this; but this kind of reduces the confidence on the testing data performance and generalizability beyond the current testing data. Any thoughts on this? thanks

[@welcomeaioverlords]

I think people can and do use the test data to improve their scores. I agree this brings up a lot of questions if you were going to use this sort of approach in a real world setting. But it depends.

[@vikramrs4191]

Thankyou for simple, lucid and articulate explanation. can you share the link of code on github as told you in the video. Thanks in advance

[@welcomeaioverlords]

Hi Vikram--glad you liked the video. All of the links are in the video's description.