Regression Probability Distribution & Multi-Quantile Output API #28060
Comments
|
To answer my own questions:
Multioutput Quantiles Distribution Parameters regressor = XGBD().fit(X, y) ) # default normal dist
regressor.predict(X) # point prediction
regressor.predict_quantile(X, quantile=[0.05, 0.5, 0.95]) # 3 x n
regressor.predict_dist_params(X) # 2 x n as mean/loc, std/scaleregressor = MultiOutputQuantile(QuantileRegressor, quantile=[0.05, 0.5, 0.95]).fit(X, y) ) # single output regressor
regressor.predict(X) # 3 x n
regressor.predict_quantile(X) # 3 x n
regressor.predict_dist_params(X) # raises error |
|
I am going to close this issue in the favor of #23334. @joshdunnlime Could you bring all your discussion on the other issue. This topic is part of the roadmap but we need to settle in some ways. |
|
I'll add any additional information from this to the previous issue. Thanks for checking back in |
Describe the workflow you want to enable
Scikit-learn has a
predictandpredict_probamethod for Classification classes but only apredictmethod for regression, with the option of quantile. Scikit-learn is adding more quantile output functionality HistGradientBoostingRegressor and QuantileRegressor - no doubt more will come in due course. The singlequantileparameter is set at the class init step.LightGBM and other packages also follow a similar API.
MAPIE allows alpha being set on class init and predict.
XGBoost also has this option but also allows multiple outputs with e.g.
alpha=np.array([0.05, 0.5, 0.95]). Currently, this isn't documented in the scikit-learn documentation. Clearly, this is a far superior piece of functionality where possible.Additionally, distributional regression packages like:
XGBoostLSS allow options on the
predictmethod such as:pred_type=quantiles,parameters,expectiles. This returns a m x n array.PGBM uses
predictwith just mean and anreturn_std=Trueoption as a 1 x n or 2 x n array.XGBD returns the mean and std as a namedtuple.
NGBoost has
predictandpred_distwhich return point predictions and the distribution parameters that can be passed to a scipy.stats distribution object. E.g.normal.All of these packages use scikit learn style APIs or aim to add this as a feature.
Describe your proposed solution
All this is to say, I think scikit-learn has the authority and opportunity to lead the way on unifying an API to cover both distributional outputs and quantile outputs. Therefore, I would like to open a discussion with the following points:
Should this be added to the core scikit learn API/package? Is this within scope?
If not, should it be an sckit-learn contrib package or is this something the "distributional python ML community" should sort out themselves?
If this is something scikit-learn would like to take a lead on, what should this API look like? If scikit-learn admin/owners think this is outside of scope but still have insights/opinions on this it would still be extremely valuable to hear them!
Describe alternatives you've considered, if relevant
No response
Additional context
No response
The text was updated successfully, but these errors were encountered: