How to handle S@98 estimation when building an honest forest

[adam2392]

1. Rejection bootstrap sampling: if a bootstrap sample does not have enough control samples (e.g. 50 for S@98) to estimate S@98 properly, then reject this bootstrap sampled indices and repeat
2. Upweight the sample weights based on class: this is the strategy sklearn currently has
3. Stratify bootstrap sample:

[adam2392]

My inclination is just do 1

[jovo]

Our estimate is broken, as evidenced by the linear simulation and more trees makes the accuracy worse. So, let's fix that. Is that a new issue, or this issue?