search

ragulpr / wtte-rnn

WTTE-RNN a framework for churn and time to event prediction
MIT License
765 stars 187 forks source link

Weibull parameters as y_train? #22

Closed NataliaVConnolly closed 7 years ago

NataliaVConnolly commented 7 years ago

Hi Egil,

I was just wondering if there are any advantages to finding the Weibull parameters through an RNN activation layer vs. first fitting the data to the Weibull to get the (alpha, beta) params and giving those to the RNN loss function instead of TTE and censored indicators? So basically separating out (alpha, beta) fitting and optimization instead of doing then jointly in the RNN. What do you think?

ragulpr commented 7 years ago

Hi there, I've been thinking about something similar. There's a big problem with having the correct initialization parameters of the output layer, but it also needs to be simple and training twice feels hacky. At present I think it's easier to initialize the output layer to be reasonable.

It will then typically figure out the rest in the few first iterations. Here reasonable is thought up as when we fix beta to 1 and alpha around mean expected tte.

Did you have problems with initialization?

Also check out test_output_lambda_initialization() in tests/test_keras.py