Franz Király

Addendum: I think only number 3 is compatible with your "reduction" style interface suggestion. I don't see how numbers 1 and 2 would map onto that, so maybe you need...

>Remove composition from crank to distr Why? I can imagine cases where this makes sense! The most obvious case would be calibration methods, e.g.: stick your predicted rank into a...

ah, makes sense.

stupid question - what precisely is the "discrete survival" task for you? Is there a summary/description somewhere?

@adibender, I know it's sth like this, but I want the precise conceptual model behind it. You can make multiple choices in discretizing: * predicting individual bins that are non-overlapping,...

>https://www.springer.com/gp/book/9783319281568 @raphaelS1, you linked an entire book, this does not answer my question clearly.

>Results make sense to me, although we might think about what featureless means in this context. @adibender, in the distribution prediction context, featureless predictions = distribution/density estimation. For example, Kaplan-Meier...

@adibender, @RaphaelS1, I feel that your conceptual model is a bit vague here. * what is your model of the tasks or scientific types? Which ones are there? * in...

Just to update this issue with information from off-line discussion with @RaphaelS1: The formula ```se(L) = sqrt(sum(cov(times))/(NM^2)``` is a plug-in estimate for the standard error of the expected generalization loss...

@RaphaelS1, @aintoha, why is this doing a massive, inelegant dispatch over existing distributions? This is not very extensible. The "natural" way would be to decorate the distribution with the squared...