Add positive argument to Ridge regression

Describe the workflow you want to enable

I’d like to have

reg = Ridge(positive=True)
reg.fit(X, y)

print(reg.coef_)

and only get non-negative print results. Adding positive to Ridge was proposed in https://github.com/scikit-learn/scikit-learn/issues/19615#issuecomment-808112174.

Describe your proposed solution

Either use scipy.optimize.nnls, which is used in LinearRegression, or scipy.optimize.lsq_linear, which has support for sparse input and LinearOperator.

Describe alternatives you’ve considered, if relevant

As the first step of lsq_linear is to solve the unconstrained least square problem, one might consider using the present ridge solvers for that and then call trf_linear or bvls.

Additional context

Linear models that already have a positive option are:

• LinearRegression
• ElasticNet
• ElasticNetCV
• Lasso
• LassoCV
• LassoLars
• …

[lorentzenchr]

I’d start with a LinearOperator and have dense and sparse support with a single minimal implementation.

[tnakae]

@lorentzenchr I opened PR #20231 to add positive argument to Ridge. This feature is often needed in my job when I have to calculate positive economical impact. I implemented it because It seems that nobody has not implemented yet.

Based on your idea, I used lsq_linear and LinearOperator, so sparse input is accepted even if positive=True. I needed to add new solver as "trf" when using lsq_linear.

Would you please check my implementation and whether it satisfies your requirement? Thank you!