LoginSign Up
← Back to all sparks
L

lime

ANALYTICS
Velocity0.0

Local interpretable model-agnostic explanations for black box predictive models

github.com/thomasp85/lime ↗

lime survives on compatibility patches years after its research moment

interpretabilitymachine-learningr-statsmaintenancexai
◆Current state
lime brings local interpretable model-agnostic explanations to R. Its substantive development finished around 0.5.0 in 2019, which added argument pass-through to predict(), a gower_pow tuning knob and a batch of fixes. Since then there have been three releases: a namespace fix, a maintainer handover to Emil Hvitfeldt with general upkeep, and a patch to work across xgboost versions.
◆Where it's heading
The package is in custodial maintenance — kept installable and compatible with the model packages it explains, rather than developed. The 2022 handover is the most consequential entry in the window because it determined that the package would keep getting patches at all.
◆Prediction
Expect the next release to be another compatibility fix triggered by an upstream model package, not new explanation methods.

◆Recent moves

  1. 8mo ago

    Compatibility across all xgboost versions

    A one-line compatibility fix so explanations work regardless of installed xgboost version. The only release in three years, and it adds nothing.

    View source ↗
  2. 3y ago

    Maintainer handover to Emil Hvitfeldt

    No user-facing change, but the entry that explains why the package still receives patches at all. Everything after this point is custodial upkeep by the new maintainer.

    View source ↗
  3. 5y ago

    order() fix and lighter dependencies

    Fixes an order() misuse and moves the Shiny-related packages to Suggests, trimming the install footprint.

    View source ↗
  4. 6y ago

    Namespace fix following glmnet changes

    A single import fix after upstream glmnet changes. Compatibility only.

    View source ↗
  5. 7y ago

    explain() gains pass-through args and gower_pow tuning

    The last release with real content: explain() forwards arguments to the underlying predict() method, gower_pow lets users reshape the distance weighting, and the default palette changes. After this the package goes quiet.

    View source ↗
  6. 8y ago

    h2o support, NA handling and date feature types

    An early broadening release: h2o models supported, missing values allowed in both training and explaining data, and date/time features handled. Shows the package in its growth phase, which the later entries no longer resemble.

    View source ↗