Description
Is your feature request related to a problem? Please describe.
You’re doing a lot of great work, but I think better OLS support would be really valuable.
I know that your focus is industry-facing advanced ML, but it's hard to introduce the framework for those advanced models in an educational setting when ML.NET cannot handle basic OLS. I'd also guess that facilitating the use of ML.NET in an educational setting is helpful for meeting long-term performance objectives such as increasing industry adoption.
More specifically, it is frustrating to try using ML.NET in an educational setting. You have great advanced features, but OLS is the standard "baseline" trainer and it's impossible to use OLS with ML.NET for teaching because the OLS trainer in ML.NET relies on MKL and is thus not cross-platform. Accord.NET, MathNet.Numerics, and FSharp.Stats all have cross-platform OLS trainers. This is entirely feasible in managed code.
Related to below issues
- Microsoft.ML.TimeSeries: Unable to load shared library 'MklImports' or one of its dependencies. #5788 (comment)
- OLS regression outputs wrong TStats and PValue #5696
- Add options for different covariance matrix estimators for OlsModelParameters #5787
Describe the solution you'd like
I should be able to use an OLS trainer on windows, linux, macOS x64 and macOS apple silicon, and all dependencies should come via (possibly platform specific) nuget feeds (no homebrew!). DiffSharp and TorchSharp are good prior art of high performance platform-specific nuget packages.