loss computation about porpoise

[Richarizardd]

What do you mean by reasonable? I think you need to distribute (1-alpha) over the other terms in Line 2.

[hachikong]

Hi @Richarizardd , thanks a lot for your kind reply!

loss 
= (1 - alpha) * neg_l + alpha * uncensored_loss
= (1 - alpha) * (censored_loss + uncensored_loss) + alpha * uncensored_loss
= (1 - alpha) * censored_loss + (1 - alpha) * uncensored_loss + alpha * uncensored_loss
= (1 - alpha) * censored_loss + (1 - alpha + alpha) * uncensored_loss
= (1 - alpha) * censored_loss + uncensored_loss

So, it seems the alpha is not applied to the uncensored_loss. I'm not sure if this is reasonable, as in the paper, the alpha should be applied to uncensored_loss.

[Richarizardd]

Hi @hachikong - Our implementation follows that of Zadeh and Schmid, which we believe is reasonable. You can see the negative log-likelihood term described in Equation 6, with the 2nd loss term (Equation 14) described at the contribution of uncensored instances in the negative likelihood.

[hachikong]

Hi @Richarizardd , thanks a lot for your kind clarification! Yeah, it seems your implementation is different from theirs'. In your implementation, the network output is the hazard function, while their network output is the survival function. Both losses are the negative log-likelihood. I will read the paper more carefully. Thank you!