How do we prevent people from overfitting on the test data by performing many submissions?
Maybe the leaderboard should initially only show the performance on the training data and only after submission is close show the performance on the real data?
How do we prevent people from overfitting on the test data by performing many submissions?
Maybe the leaderboard should initially only show the performance on the training data and only after submission is close show the performance on the real data?