Adding LightGBM

Is your feature request related to a current problem? Please describe. LightGBM won the M5 time series forecasting competition. It will be a good algorithm to add. https://www.innovating-automation.blog/m5-forecasting-competition/

Describe proposed solution Create a wrapper for the lightgbm library in darts

[hrzn]

We have released the changes explained by @tneuer above in version 0.8.0. So now you can

• Wrap around any scikit-learn model
• Wrap around a LightGBM model (offering fit() and predict() interface) if you install LightGBM yourself on your environement.

I’ll thus close the issue, but we would be happy to learn about your experiences using these models!

[tneuer]

Hey @RNogales94, thanks for bringing up the issue again. We have also been working on this problem and a few general notes to update all interested folk:

• You are right that Tree-Based-Models tend to predict the mean (that’s the operation performed on the leaf-level) and sometimes have troubles with time series, especially when you deal with a time series with non-zero trend (like the passenger data set). The user should perform differencing before passing any TimeSeries to a Tree-Based-Model (this feature has just be enabled in Darts by calling .diff() on any TimeSeries.)
• For now, we decided not to implement the LightGBM class from the lightgbm library, because it requires additional dependencies outside of python which cannot be downloaded via conda. On a Mac you need to perform these steps to make lightgbm work and we already have so many Python dependencies that we decided against having even more out-of-Python dependencies which would break the Darts installation.
• In the near future we release models wrapping around Random Forest and HistGradientBoostingRegressor from scikit-learn (it is currently in the PR and we need to review it properly).
• In the same release we enable the user to use Darts for every model exposing the .fit() and .predict() methods (so every scikit-learn model). You then could download LightGBM separately and then pass that model directly to the base class called (for now) RegressionModel to make everything work.

@RNogales94 I would still be eager to see how you implemented the model. Maybe we see something that would improve on our implementation. As it is still in the PR we could change and improve it without causing any problems. Would you mind making a PR for your implementation. It will probably not be accepted due to the mentioned concerns with lightgbm but I think that would still be the easiest way to review the code together.