LoginSign Up
← Back to all sparks
D

DMRnet

INFRA · APIS
Velocity0.0

Delete or Merge Regressors Algorithms for Linear and Logistic Model Selection and High-Dimensional Data

github.com/szymonnowakowski/dmrnet ↗

A categorical-variable selection package that publishes its full test logs as release candidates.

variable-selectionhigh-dimensionalcategorical-datar-packagecross-validationstatistics
◆Current state
DMRnet implements delete-or-merge-regressors model selection for high-dimensional categorical data, alongside SOSnet and GLAMER variants from the same research group. Development is slow and academic — 0.4.0 in 2023, then two years to 0.4.1 in August 2025, which corrects an invalid lambda.1se computation and the cross-validation plots that displayed it. Every real release is preceded days earlier by a release-candidate entry containing the raw output of the correctness and consistency test suite.
◆Where it's heading
The package is converging on correctness rather than expanding. 0.3.3 was a wall of fixes to inference, log-likelihood, and degenerate cross-validation cases; 0.4.0 added the var_sel algorithm and brought GLAMER into the package's own net idiom over its tau parameter; 0.4.1 is again a statistical correctness fix. The published test-log releases are the tell — this maintainer treats reproducible evidence that hard cases still pass as part of the release artifact, which is unusual outside academic statistical software.
◆Prediction
Given the two-year gap before 0.4.1 and its narrow scope, the next release is most likely another correctness fix arriving on a multi-year cadence, again preceded by a full test-log release candidate.

◆Recent moves

  1. 1y ago

    Invalid lambda.1se computation corrected in cross-validation

    A statistical correctness fix rather than a cosmetic one: lambda.1se was being computed invalidly, which means cross-validated model choice was wrong, and the CV plots were redrawn to match the corrected values. Continues the pattern from 0.3.3 of hunting down errors in the selection machinery itself.

    View source ↗
  2. 1y ago

    Test-suite log published ahead of the 0.4.1 release

    A release-candidate entry containing the raw stdout of the correctness and consistency suites — airbnb, insurance, promoter, and simulation hard cases — rather than a changelog. It is evidence attached to the 0.4.1 release five days later, not a shipped change.

    View source ↗
  3. 3y ago

    var_sel added; GLAMER reworked as a net over tau

    The last feature release before the two-year gap. var_sel adds another variable-selection routine, GLAMER is reimplemented as a net over its tau parameter so it fits the package's existing path idiom, and hierarchical clustering moves to the hclust1d package for exact segmentation of one-dimensional points in PDMR and GLAMER.

    View source ↗
  4. 3y ago

    Test-suite log published ahead of the 0.4.0 release

    The 0.4.0 release candidate, again a dump of hard-case and massive-consistency test output rather than release notes. Same publish-the-evidence practice that precedes every DMRnet release.

    View source ↗
  5. 3y ago

    Inference, log-likelihood, and degenerate-CV fixes across all model families

    A dense correctness release. Inference was wrong when columns were not ordered factors-then-numerics, log-likelihood was wrong or -Inf for the largest binomial model, and cross-validation broke when few or no models survived. The no-model fixes were then propagated from gaussian DMRnet out to the binomial family, SOSnet, and GLAMER.

    View source ↗
  6. 3y ago

    Test-suite log published ahead of the 0.3.3 release

    The 0.3.3 release candidate — correctness and consistency test output published two days before the release itself. No user-facing change.

    View source ↗