on log likelihood misunderstanding about ifl-tpp HOT 4 CLOSED

You are correct that log_prob() computes log-intensity plus log survival probability (i.e., log intensity minus the integral). In the code (lines 128-130) I only compute the log survival probability for the time interval [t_N, T), i.e. between the last event and the end of the observed time interval. Does this answer your question?

from ifl-tpp.

Thanks. However, in lots of implementation, the final time interval survival probability is omitted, and do you agree that this can be all omitted, which will still be a strict model comparison.

from ifl-tpp.

I would be very careful when comparing to other implementations. Small changes (like rescaling / normalizing the times) can totally change the log-likelihood and make the results incomparable. I have also seen some implementations that compute the loss differently, like only computing the log-likelihood for marks, adding some regularization terms, or sometime even having serious bugs in the computation.

from ifl-tpp.

Thanks a lot. I will continue following and expanding your fantastic work!

from ifl-tpp.