Multivariate predict univariate

[Alexey322]

Hello. As I understand from your paper and discussions in the repository, the model is trained to predict each time series independently of other time series.

So, let's say if we have three parameters: age, sex, salary. Then, training your model is equivalent to training three neural networks that are trained to predict univariate age based only on previous age values, univariate gender based only on previous gender values, and univariate salary based only on previous salary values.

The "--features" parameter in PatchTST_supervised/run_longExp.py says that you can select MS - multivariate predict univariate. By using the MS parameter instead of calculating the error of all time series, it only calculates the error for the target time series.

My question is why is it called "multivariate predict univariate" instead of "univariate predict univariate" and what is the point of using other input time series if they have no effect on the target time series?

I will be grateful for a detailed answer. @yuqinie98 @namctin

[chendingyan]

Still confused with this

[alceubissoto]

I think it's because even though each time series is treated independently of others, they all share the same backbone, inherently influencing eachother.