Work in Progress
A tiny ridge regression library for multiple targets, with optional automatic selection of the ridge parameter (Using the estimator ‘k36’ from this paper (see references), for all, not only Poisson regression cases. )
Obviously this is highly experimental and has no mathematical guarantees… but I try to have a curated set of benchmarks to showcase whether it works in practice.
The goal here is absolutely NOT to provide an interface like glm or glmnet, but rather to have a lean, easy to understand set of solvers, with models capable of producing predictions. These solvers should be easy to plug-in where-ever you need to solve a ridge regression problem, and not much else.
It also serves as a simple educational tool - since the implementation is primarily concerned with simplicity and does no unnecessary calculations, it can be used to reasonably showcase ridge regression in a generalized linear model context. While elastic-net could be considered a better approach in many contexts, I believe most implementations are opaque and hard to understand.
To the goal of simplicity, this package currently supports no scaling/re-scaling, no addition of intercepts, no formulas, and absolutely no offsets. The scaling and intercepts might be added at some point, but I would prefer to keep this contained as a very bare-bones implementation.
Currently only from github.