0
$\begingroup$

as you can see in the image below, I need to bridge static data (on the left) and time series data (on the right) to create a time series output. I have looked at this example on keras library which deals with transformers but to no avail. generally what are the layers that are used to solve:

time_series + static -> time_series

would the "core" of the answer change if it were (=core maybe you keep the same framework and then you just pick the latest value or "reduction" of some kind): time_series + static -> static

enter image description here

here the time series part is more "features" based on time that changes and I would avoid using RNN/LSTM or the sort as much as possible. it's just a matter of concatenating things without resulting in not exploding my features and therefore adding "bias" to the problem

p.s.: I know there are some custom solutions for this problem but here I need to deal with transforming/concatenating things as I will add multiple sources like this (this is a simple example) and I need a general way to deal with this kind of aggregation problem (feeding everything to a single RNN will not do)

Improve this question
$\endgroup$
5
  • $\begingroup$ Why an algorithm like random forest would'nt solve your problem? Does an adapted data pre processing to scale the features correctly between each other would be the key? $\endgroup$
    – Nicolas Martin
    Commented Feb 2, 2022 at 21:17
  • $\begingroup$ because I might have multiple data sources that would generate some multiplication of features {users} X {products} X {time_series_users} X {time_series_products}X {time_series_other} -> ideally I would crunch this in a smart fashion $\endgroup$
    – Asher11
    Commented Feb 3, 2022 at 8:43
  • $\begingroup$ In that case, is it possible to weight each feature so that they can be classified correctly? I mean that some features can be common to lot of data sources, the other ones can be included but with some mathematical operations to be correctly weighted between each other. $\endgroup$
    – Nicolas Martin
    Commented Feb 3, 2022 at 14:07
  • $\begingroup$ what do you mean? could you give a simple but practical example? $\endgroup$
    – Asher11
    Commented Feb 3, 2022 at 14:57
  • $\begingroup$ Sorry, I thought you had heterogeneous data sources, but no: you just have several sources that you want to group in one. In your case, it could be meaningful to make a time series model for each main feature (one for the users, and one for the products) with the right time scale (one user bought 5 products in one month). If you want to do multiple time series with different times features, it is also possible but maybe a time simplification (ex: convert to day or weeks) could be more interesting to group data better. $\endgroup$
    – Nicolas Martin
    Commented Feb 3, 2022 at 15:17

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.