WSDM2019-nextitnet
WSDM2019-nextitnet copied to clipboard
loss needs masking for padding labels?
Hi Fajie, I have a question regarding the loss in https://github.com/fajieyuan/nextitnet/blob/master/generator_recsys.py
users' histories have different length, so padding zeros will be needed in input data. And for the loss calcutaion, we need to skip those padding parts, right? But it seems the code doesn't take that into account. Any thoughts about this? Thanks in advance.
Hi it has no impact, you can also mask zero padding which is faster but the accuracy should be more or less the same