Replies: 1 comment
-
Hi @mputra533, thank you for the question! All forecasting models in AutoGluon are univariate, which means that each model generates predictions for each time series in "isolation", without looking at other time series. In other words, after we trained the model, if we change history of item A this won't affect the forecast of item B. However, some models (such as DeepAR or WeightedEnsemble) are trained globally. This means, during training we are looking for a single model that on average performs best for all items. In case of ensemble, this means that the ensemble weights are shared across all items. This means that in your example:
|
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi!
A few questions about training multiple items in one big dataframe (separated by item_id)
How does training for multiple time series/item work?
Does the time_limit apply to each model for all items or each model for one item?
Documentation states "AutoGluon generates forecasts for each time series individually, without modelling interactions between different items" Wondering why there's only one best model and validation score for all of the multiple items that are trained as one from the leader board?
I did an experiment where I have 4 normal items and one dummy item with all 0 values. I trained all 5 items at once from one dataframe, separated by the unique item_id's. I expect the item with all 0 values to have prediction of all 0's but it wasn't. The prediction seems to be influenced by the other 4 normal items? I sort of followed the trend from the other 4. But, when I train the item with all 0 by itself, it predicted all 0's - which is what I expected.
Beta Was this translation helpful? Give feedback.
All reactions