i agree with this question, pls. answer, organizers! i have been very hesitant to do any type of class weighting as it appears to be prohibitied by the rules - but there is no way to tell if leaderboard submissions are using it!

Hi @meganomaly / @AJoel,

Follow-up to the earlier "Threshold clarification" thread — I want to check a specific technique against the rule that raw probabilities must reflect the default 0.5 threshold, with no threshold-setting or rounding to improve leaderboard placement.

You've confirmed the test set's positive rate is inherently higher than the ~40% train rate. Given that, is it acceptable to apply a prior-probability-shift correction (Saerens, Latinne & Decaestecker, 2002) to our raw model outputs before submission?Concretely: we estimate the test-set positive rate from our own out-of-fold TPR/FPR bias (Adjusted Classify & Count), then apply the standard closed-form Bayesian correction to each row's probability, given the known train prior and the estimated test prior:corrected_p = (test_prior/train_prior * p) / (test_prior/train_prior * p + (1-test_prior)/(1-train_prior) * (1-p))

This produces a genuinely updated per-row probability estimate reflecting the known prevalence shift you've described — not a uniform quantile remap chosen to hit a target leaderboard score. TargetRAUC would be this corrected probability, and TargetF1 would be (corrected_p >= 0.5) at the standard threshold. Is this within the rules, or does it fall under the prohibited threshold/probability-adjustment behavior? A clear yes/no would help everyone avoid either under- or over-correcting for the prevalence difference you've flagged.

Thanks!