Re:Kaggle Post-Mortem: The dangers of overfitting
by chris_wot on Saturday July 07, 2012 @09:10AM
(#4211544)
Why even look at these scoreboards?
[ Parent
| on HN ]
Re:Kaggle Post-Mortem: The dangers of overfitting
by shoo on Saturday July 07, 2012 @10:45AM
(#4211736)
Assume that the validation data is similar to the test data. Then, by performing experiments (making submissions) and recording measurements (analysing the returned leaderboard score) you can attempt to infer something about the structure of the validation data, and make more accurate predictions about the test data. You can probably infer quite a bit of useful information by looking at other people's scores too, even if you don't have access to their submitted predictions.