What sort of model should I use

Last updated Feb 27, 2024 Edit Source

#work/patientsim
2024-02-07

Well, there’s lots of examples. I’m pretty sold on those triplets because they’re cool and they solve the whole time+feature space sparsity issue in a pretty way. That’s really more of a way of fitting the data in. Definitely superior to how it’s done in the others, though. They’ll:

• cut it into windows, create overlapping embeddings
• give each feature their own entire NN
• stuff it into an RNN and just grab the whole final output, throw that into the main solution
  • or worse, they’ll aggregate the RNN output by summing over the outputs of each step???? @merkelbach
  • at least one of them computes the attention or smth over the outputs of each step. idk which tho

So, it comes down to what I’m piping the triplets into. I could really pick anything. There’s really two routes:

• wild - try out the $CNN$ for time series
• basic - just do another transformer™

my choice