LoginSign Up
← Back to all sparks
A

abclass

ANALYTICS
Velocity0.0

R package abclass by wenjie2wang — release notes from GitHub.

github.com/wenjie2wang/abclass ↗

abclass built out angle-based classifiers in 2022, then went quiet except for CRAN upkeep.

classificationregularizationlarge-margin classifierscran maintenancer package
◆Current state
An implementation of multi-category angle-based large-margin classifiers with regularization. The capability was assembled in four releases across 2022: group lasso, then group SCAD and MCP penalties, then sparse matrix input, cross-validation via cv.abclass(), an efficient tuning path in et.abclass(), and experimental sup-norm classifiers. After a three-year gap, 0.5.0 simplified how group penalties are specified and 0.5.1 swapped the quadratic programming backend after qpmadr was archived on CRAN.
◆Where it's heading
The methods surface is complete and the package has moved into maintenance, where releases are triggered by the R ecosystem rather than by research. The one structural habit worth noting is a willingness to change defaults — alpha, epsilon, lum_c and now the cross-validation alignment have all shifted between versions, so results are not stable across upgrades unless arguments are set explicitly.
◆Prediction
Expect further releases to track CRAN dependency changes, as 0.5.1 did within a day of qpmadr's archival; nothing in the entries points to new penalty families.

◆Recent moves

  1. 7mo ago

    Quadratic programming backend swapped after CRAN archival

    Swaps qpmadr::solveqp() for quadprog::solve.QP() because qpmadr was archived on CRAN the previous day. A forced backend substitution with no intended change in behaviour.

    View source ↗
  2. 10mo ago

    Group penalty specification simplified

    Simplifies how group penalties are specified through abclass.control() and changes the default alignment to lambda for cv.abclass() and et.abclass() refits. The first functional release in three years.

    View source ↗
  3. 3y ago

    Sparse input, cross-validation and efficient tuning added

    Adds sparse matrix input, splits cross-validation out into cv.abclass() with an efficient tuning path in et.abclass(), and introduces experimental sup-norm classifiers. The release that turned a fitting function into a usable modelling workflow.

    View source ↗
  4. 4y ago

    Group SCAD and MCP penalties added

    Adds experimental group SCAD and group MCP penalties and introduces abclass.control() to carry control parameters instead of a widening argument list.

    View source ↗
  5. 4y ago

    Group lasso regularization and correctness fixes

    Adds experimental group lasso regularization along with a run of correctness fixes — boosting loss derivatives, label prediction now using fitted inner products, and validation accuracy in cross-validation. Early enough that results changed materially between versions.

    View source ↗
  6. 4y ago

    First release of the angle-based classifiers

    The first release, providing multi-category angle-based large-margin classifiers with a choice of loss functions. Everything since has been regularization and tuning built on top.

    View source ↗