LoginSign Up
← Back to all sparks
B

bolasso

INFRA · APIS
Velocity0.0

Model Consistent Lasso Estimation Through the Bootstrap

github.com/dmolitor/bolasso ↗

Bootstrap lasso got a fast mode, a second selection rule, and multinomial support

variable-selectionlassor-packagebootstrapstatisticsregularization
◆Current state
bolasso implements the bootstrapped lasso, refitting a regularized regression across bootstrap replicates and selecting variables by how consistently they survive. The 0.3.0 release reshaped it: a fast argument computes one cross-validated lambda on the full dataset instead of cross-validating inside every replicate, and selected_variables() gained a choice between the variable inclusion probability rule and a quantile rule based on bootstrap confidence intervals. Since then 0.4.0 exposed the bootstrap indices through bootstrap_samples(), and 0.5.0 extended the whole surface to multinomial responses, returning one list element per outcome level.
◆Where it's heading
The package spent 2022 dormant after its initial releases and has been actively developed since late 2024, moving from a single algorithm toward a workbench. The additions cluster around inspection rather than estimation: tidy() for bootstrap-level coefficients, plot_selection_thresholds() for selection stability across thresholds, plot_selected_variables() for the surviving covariates, and now the extracted bootstrap indices. Documented gaps remain, with mgaussian unsupported and multinomial prediction limited to class output.
◆Prediction
The two stated limitations in 0.5.0 - no mgaussian family and class-only multinomial prediction - are the most likely next targets, since the maintainer flagged both as possible later additions.

◆Recent moves

  1. 4mo ago

    Multinomial responses supported across selection and plotting

    Makes family = "multinomial" work throughout rather than partially: outputs become lists with one element per outcome level, keeping their original structure, and plots facet by outcome. The maintainer is explicit about what remains missing - mgaussian entirely, and multinomial prediction restricted to type = "class" - which is the useful kind of release note.

    View source ↗
  2. 10mo ago

    Bootstrap indices exposed via bootstrap_samples()

    Adds an accessor for the bootstrap indices behind a fitted model, letting users reproduce or audit the resampling rather than treating it as internal. Consistent with the inspection-oriented direction the package took in 0.3.0.

    View source ↗
  3. 1y ago

    Fast estimation mode and a second variable selection rule

    ⚡ SPARK

    The release that ended the package's dormancy and set its current direction. Everything after it - the bootstrap accessor, the multinomial extension - builds on the inspection surface introduced here rather than changing the estimator.

    View source ↗
  4. 4y ago

    Bolasso v0.2.0

    The release note points to NEWS.md rather than describing the changes, saying only that there are minor enhancements and bug fixes. Direction cannot be read from the feed here. It is also the last release before a two-and-a-half year gap.

    View source ↗
  5. 4y ago

    Bolasso v0.1.0

    Initial CRAN release, announced without any description of contents. What the package did at launch has to be inferred from later releases rather than from this entry.

    View source ↗