I've been trying to make a Neural Network for handling requests for quite sometime but I keep facing this bug where the predictions are not even binary. The independent variable is just the time on the clock and the target is when the request happens.
I've even tried to make the requests last longer so the system doesn't get addicted to zeros (the custom loss functions were an attempt to help solve the sparsity problem, but they don't seem to work either).
I expect the problem to be easy for someone more or less experienced, and so would be really really thankful if the final code had an LSTM architecture instead of the normal one I'm using now. I have some other Deep Learning projects I need help with also, so if you can nail what I'm doing wrong here, be sure that more jobs will come to you soon!