Thread · 2 posts
Auditing how coefficients and predicted labels vary among near-optimal GLMs. The thread develops the loss-set geometry and toolkit, then compares sampling and the quadratic approximation with exact optimization.
Posts follow their reading order. Each shows its publication date and latest revision.
Reading path
The double ring marks the latest post.
The Rashomon effect means many parameter vectors achieve nearly the same loss. StableGLM characterizes this set for GLMs and asks which feature importance claims are stable across it.
Compare against exact optimization
The geometric approximation and sampled summaries need a reference. The benchmarks compare them with exact coefficient ranges and label-flip tests, revising the earlier interpretation of sampling and interval widths.
Comparing exact coefficient ranges and label-flip tests with sampling and the Hessian approximation on three classification datasets.