Hi! I'm using MLForecast for multiple unique ids, ...
# mlforecasty
Yaarit Even
02/05/2025, 9:26 PMHi! I'm using MLForecast for multiple unique ids, and I'm trying to understand how does the model work in terms of learning between different series (I'm using the LGBM regressor). Thanks!
j
José Morales
02/05/2025, 9:28 PMthe model is trained on all series
y
Yaarit Even
02/05/2025, 9:38 PMHow does it account for the unique ID then? Since it's not a feature...
Yaarit Even
02/05/2025, 9:41 PMWhen doing inference for a series that was in the training data, will it give more weight to the learning from that specific series or is it just one model regardless of the unique id?
From the explanation online it sounds like the unique id matters.
j
José Morales
02/05/2025, 9:44 PMthe id is only used to generate the features but it can be a feature as well if you provide it through
static_featuresy
Yaarit Even
02/05/2025, 9:46 PMSo if I don't use the unique id as a static feature explicitly, once the model is trained and we are doing inference there's no importance to the unique_id?....
j
José Morales
02/05/2025, 9:46 PMit's one model, so it doesn't even know it's dealing with different series unless you provide it as a static feature, in which case it'll be used as a categorical
y
Yaarit Even
02/05/2025, 9:47 PMoh, interesting. So we don't really need the unique id once the model is trained
j
José Morales
02/05/2025, 9:48 PMyou need it for the forecasts because we generate future dates as a convenience, but it's just for that, convenience
👍 1
y
Yaarit Even
02/05/2025, 9:50 PMunderstood. Thank you for clarifying!
one last clarification then - for training purposes, the unique_id is only used to generate features, and it does not "learn more" from series with the same unique id?
j
José Morales
02/05/2025, 10:00 PMthe unique_id is only used to generate features, and it does not "learn more" from series with the same unique id?exactly. the model just sees all of the series at the same time and doesn't care about which is which
y
Yaarit Even
02/05/2025, 10:01 PMok, thanks again for the clarification!
9 Views