We should be able to do some optimisation on the mixing coefficients when we average submission csvs to optimise our score on the leaderboard. We can do this using the holdout set.
Best way to code it is probably to call the function in check_test_score that produces predictions and then input this to sklearn. Then, fit a logistic regression model to each class individually (one vs. all) in the holdout set. Check the weights that we get out doing this to make sure they're reasonable.
We should be able to do some optimisation on the mixing coefficients when we average submission csvs to optimise our score on the leaderboard. We can do this using the holdout set.
Best way to code it is probably to call the function in check_test_score that produces predictions and then input this to sklearn. Then, fit a logistic regression model to each class individually (one vs. all) in the holdout set. Check the weights that we get out doing this to make sure they're reasonable.