Pseudolikelihood estimators

[mnarayan]

sklearn.linear_model has all the ingredients to implement the "pseudolikelihood" regression version of inverse covariance estimation that doesn't have the wrapper block coordinate descent. It is literally p uncoupled high dimensional regressions with some extra steps to symmetry and positive definiteness using cholesky decompositions.

It would be easy for us to implement a pseudolikelihood estimator that does this properly using sklearn.linear_model functions as the internal. This can serve as a comparison to loglikelihood type estimators that QUIC-dirty gives us for mixed norm penalties

To be fleshed out more

References:

• Using Regression + Cholesky decomposition
• Pseudonet
• Concord

[mnarayan]

This would wrap around both sklearn's linear model estimators (for Gaussian data) and pyglmnet for non-Gaussian data https://github.com/pavanramkumar/pyglmnet/tree/master/pyglmnet.

Very usable for people working with binary and poisson observations.