In my project, for each sample (a patient) I have some features that are time-varying (weight, blood pressure) and a bunch of others that are constant (male/female…). Presently, I pass on the data for each sample to my RNN as a large matrix of data (time steps x features), much of which is redundant.
Does anybody have any suggestions on how to intelligently incorporate the time-invariant data into the network?
(My architecture is a single LSTM layer and a single dense layer on top.)
how did you end up solving this? Using Keras funcitonal API you can merge the output of the LSTM with your time-invariant data, and add one (or several) dense layers afterwards. I think somewhere in the videos (don’t ask me where) Jeremy mentions people do that when processing images (after the CNN does its thing, metafeatures are merged with the result of the latter and everything is fed to a dense layer or whatever).