Description
The Iteratively reweighted least squares (IRLS) solver, initially proposed in #9405 could be used in several linear models, for instance,
- LogisticRegression
- TwieedieRegression (including Poisson & Gamma) etc
Preliminary benchmarks for TweedieRegression were done in #9405 (comment) and demonstrated that IRLS could be competitive (or faster than other solvers including LBFGS) when n_features ≤ 10-20. More detailed benchmarks with a careful consideration of the stopping criterion would be necessary.