How to use exogenous variables in iTransformer? about itransformer

[ben10ben]

With the current implementation, I am playing with, let's say 1 target and 4 covariates. For the first couple of epochs, I use all variates for loss calculation, for the final tuning I just use the target for loss calculation. For now, I have not done a meaningful evaluation, technically it is possible at least.

[lucidrains]

@meteoDaniel yea i have my own ideas, but it would take a long research project to prove it out, with uncertain results. i would probably only do it as a contract project under calibrated expectations.

[lucidrains]

@meteoDaniel are you already seeing positive results as is? without exogenous features?

[lucidrains]

@meteoDaniel if you wait a year, i'm sure we'll see a few more papers building on this

[lucidrains]

anyways, let us move this to discussions, as it isn't an issue with the implementation of the paper

[kashif]

due to the arch. of this model, only per time series static covariates can be incorporated, namely static ids per time series which can be passed through embedding layers and the resulting vector concat to the representation of each time series, or real-valued static covariates that can be similarly concated. If those covariates have predictive information then the results should be better with them. time-varying covariates would need to be dealt with differently and it's not as straight-forward...