score_ -> log_likelihood

[CamDavidsonPilon]

I'm increasing like using the log-likelihood as a measure of "fit" in sample and out of sample.

At the moment, a model's score_ is the c-index, which I'm increasingly not liking.

I may consider swapping these two...

[KSafran]

IMHO this would be a good change.

Related question - why just do log-likelihood for the uncensored data. Couldn't you also calculate a likelihood for censored data using 1 - CDF? Isn't log-likelihood for uncensored points only biasing the predictions downward (due to sampling bias)?

[CamDavidsonPilon]

@KSafran, can you expand a bit more on your question? I am not understanding where in the code the likelihood for censored is not calculated.

[KSafran]

Ah, my bad, I was looking at the log loss metrics functions defined here, but looking at these loss functions it looks like you do account for the likelihood of right-censored values.