What does the LSTM predict? It looks like the survival rate, not the true event time.

[jialin666]

Hi there; thanks for your contribution and for sharing the code. When I review the code, I get stuck in some places. First, It looks like you treat the output of the model as the survival rate, not the probability of true event time (BASE_MODEL, line 299). Second, what confused me most is how you dealt with the output, in the function reduce_mul, the code generated production of the output from time step 0 to time step MAX_SEQ_LEN! Why is MAX_SEQ_LEN? I thought it should be the time step that the event occurred. Could you help me to resolve these confusion? Thanks!

[jialin666]

Everyone is very welcome to discuss the code of this paper with me ([email protected]).

[rk2900]

Hi,

1. According to our paper, the hazard rate would be used for calculating the survival probability. Please read the details of our paper.
2. Your understanding is correct. The calculation are based on the whole sequence with max sequence length. But, the results of interest at the event time are gathered later in Line 305. This is to fully utilize the parallel computation of GPU and some survival probability calculation (after the event time) would be regarded.