grobgl / rnnsm Goto Github PK

Hi Georg,

First, thanks for making your code available! It's a really interesting approach to survival analysis.

Can I ask you a couple of questions?

1. How do I generate the data? All the notebooks seem to assume that there's a directory called "data" with a bunch of .pkl files in it, but the directory is not there on your github.

2. Is there a way to estimate feature importance in this approach?

   Thank you!

Hi Georg,

I've greatly enjoyed reading your paper and appreciate you making the code available!

Could you please help me understand Eqn. (21) in your paper and its implementation in the code? Eqn. (20) gives the expression for the conditional expected return time E[T|T>t_p] where t_p = start of prediction window. But then in Eqn. (21) t_p gets replaced by t_s, and I am not sure what that is. Looking at the code in rmtpp.py, it looks like this def corresponds to Eqn. (21), correct?

def pred_next_starttime_rec(self, cur_state, t_j):
    absence_time = 365*self.time_scale - t_j
    s_ts = self._pt(absence_time, cur_state)

    ts = np.arange(t_j, 1000*self.time_scale, self.time_scale)
    delta_ts = ts - t_j
    samples = self._pt(delta_ts, cur_state)

    return (1/s_ts) * trapz(samples[ts>(365*self.time_scale)], ts[ts>(365*self.time_scale)]) + trapz(samples[ts<=(365*self.time_scale)], ts[ts<=(365*self.time_scale)])

So then it looks like t_s is 365 days. I am not sure what that means? Doesn't this imply that we are requiring return times to be after a year? If so, why? Why not use predPeriod[start] for t_s?

I would greatly appreciate your help with this!

Thanks so much,
Natalia

I read the great paper and was eager to try out the model. Having the data available would be helpful. Any plans on posting the datasets or references to where they can be found?