loss function focusing on important errors

[lyeskhalil]

Idea: Setting a low bias threshold for binary classification seems to lead to "better" training. However, this comes off as "hacking" the labels.

Proposal: instead, do regression with an appropriate loss function. The loss function should incur higher cost for underestimating the bias of a variable with true bias non-zero but close to zero.

[lyeskhalil]

This seems like an option: https://datascience.stackexchange.com/a/10474

[chrsmrrs]

Talked to Didier about the [0,1] issue, he mentioned doing a logit transformation to map them to the real interval or using beta regression, see, e.g., https://stats.stackexchange.com/questions/29038/regression-for-an-outcome-ratio-or-fraction-between-0-and-1.