Training a model to forecast a TS for several subjects.

[darleybarreto]

Hi folks, first I would like to say great job with FEDOT, it certainly is an incredible piece of software! For what I understood by looking at the examples, I can only train a model (or an automl pipeline + hyperopt) for each TS. So if there's several TS, one for each subject of the dataset, I would need to train one model for each one. Is that correct? Or there's a way of incrementally training, i.e., train for one subject, than the next, the next, and so on.

Thanks in advance!

[nicl-nno]

Hi, Darley.

Thank you for such positive feedback!

The described problem of incrementally training is similar to a hierarchical time series forecasting task. It is not implemented yet, but we already have several requests to add it. Conceptually, it fully satisfies the ideas behind FEDOT.

However, there is ambiguity in practical implementation (structure of input data, etc). It will be very useful if you can provide us the script with an example of the solution of described task using an approach that you noted as "need to train one model for each TS".

It will help us to design a fully FEDOT-based solution to it in the future.

[Dreamlone]

Would the model understand which rows correspond to which devices and learn trends and seasonality?
Yes, if you place the row values opposite the value of the month and the device.

Regarding seasonality and trending - just try running the model. If you do not get a satisfactory result, you can count the characteristics of the components of the time series, or get values after the Fourier transform.

But first I advise you to try to build a model on your source features

[darleybarreto]

I will look into that. Thank you so much for your help!

[darleybarreto]

@Dreamlone I looked at this classification example. If the data are in this form, is automl a nice fit to find a pipeline? Or should I use the pipeline tuning?

[Dreamlone]

This pipeline tuning example shows how you can create your own pipeline without AutoML, and then tune it :)

Yes, the example works for any tabular data and you can use it for yours

However, keep in mind that pipeline tuning is part of the AutoML algorithm. That is, when you run AutoML (as here, you can check logs for phrase Hyperparameters optimization start) the model search goes through two steps:

• composing (finding the optimal topology)
• tuning (tuning hyperparameters)

That is, you don't need to use tuning after applying AutoML. The found Pipeline has already gone through this procedure.

[darleybarreto]

Since automl-ing my train protocol takes lots of hours, what I had in mind was using automl to find a "decent" pipeline and then specify that pipeline by hand to train from scratch for more time, in this case also tuning it. What I mean by decent is making the timeout really small, so it probably wouldn't find the best one.