[dask][docs] initial setup for Dask docs

[StrikerRUS]

Towards #3814.

Maybe we can ask to archive old dask-lightgbm repo?

Live demo: https://lightgbm.readthedocs.io/en/dask_docs/Python-API.html#dask-api.

Why don't we have the most general class DaskLGBMModel?

[jameslamb]

Maybe we can ask to archive old dask-lightgbm repo?

After LightGBM 3.2.0 is released, I'd like to make a pull request in dask-lightgbm that adds a warning to all public functions and constructors saying "dask-lightgbm will not receive further updates. Please use lightgbm instead ('pip install lightgbm[dask]')". If they agree, then after that release we can ask for that repo to be archived.

Why don't we have the most general class DaskLGBMModel?

What is the value of LGBMModel? I don't understand why someone would use that instead of one of the task-specific classes (DaskLGBMRanker, DaskLGBMRegressor, DaskLGBMClassifier). Could you help me understand that?

[jameslamb]

thank you for this! I left some small suggestions, but I agree with these changes.

[jameslamb]

Suggested change

	raise LightGBMError('Dask, Pandas and Scikit-learn are required for this module')
	raise LightGBMError('dask, pandas and scikit-learn are required for lightgbm.dask')

Instead of "this module", could you use the specific name? I think that makes the log message a little more useful standalone. It can be helpful for cases where people don't have direct access to the stack trace, which is required to understand what "this module" refers to.

For example, user code or other frameworks might write things like this

try:
    dask_reg = DaskLGBMClassifier()
except LightGBMError as err:
    log.fatal(err)
    raise SomeOtherException("LightGBM training failed")

I also think packages should be referenced by their exact package names, not capitalized names.

[StrikerRUS]

Thank you! Addressed in acac78f.

[jameslamb]

oh wow, thank you!

[StrikerRUS]

TBH, first time I noticed that was LGTM site:
https://lgtm.com/projects/g/microsoft/LightGBM?mode=tree

[StrikerRUS]

@jameslamb

After LightGBM 3.2.0 is released, I'd like to make a pull request in

Great!

What is the value of LGBMModel?

I believe this class is needed for extending scikit-learn features by supporting more objectives. For example, right now LightGBM supports cross-entropy application. LGBMClassifier cannot be used with this objective because scikit-learn checks targets to be ints

LightGBM/python-package/lightgbm/sklearn.py

Line 818 in ac706e1

_LGBMCheckClassificationTargets(y)

One can workaround this with LGBMRegressor and cross_entropy objective but I don't think this is semantically correct.

Some other checks can be added in RegressorMixin and ClassifierMixin at scikit-learn side that will prevent using some potential new LightGBM features.

So in general LGBMModel is for everything beyond classification and regression.

[jameslamb]

these changes look good, thanks very much!

[jameslamb]

So in general LGBMModel is for everything beyond classification and regression.

Ok, thanks for the explanation! I've created #3845 for the feature request.

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.