LoginSign Up
← Back to all sparks
X

xplainfi

ANALYTICS
Velocity2.5

Model-agnostic feature importance methods

mlr-org.github.io/xplainfi ↗

xplainfi treats feature importance as an estimate with error bars, not a number.

mlr3feature-importanceinterpretabilitystatistical-inferenceconfidence-intervals
◆Current state
xplainfi implements feature importance methods for mlr3 — perturbation-based PFI, CFI and RFI, refit-based LOCO and WVIM, and SAGE. Its defining choice is that importance scores come with inference attached: several confidence-interval methods, including the Nadeau-Bengio correction and a distribution-free option added in 1.1.0. It declared itself released at 1.0.0 in January 2026.
◆Where it's heading
Two lines of work run in parallel. The statistical side keeps adding inference options — variance corrections, conditional predictive impact, and the Lei et al. observation-wise loss-difference test — while the computational side attacks the cost of refit-based methods, most recently with a batch_size argument that parallelises refits and a default of one refit per resampling iteration. Support for pre-trained learners in 1.1.0 removes the refit requirement entirely in some workflows.
◆Prediction
The stated reasoning that budget is better spent on resampling iterations than repeated refits suggests n_repeats may be removed from WVIM and LOCO outright, as the release notes hint.

◆Recent moves

  1. 20d ago

    Refits parallelise; repeated refits deprioritised in favour of resampling

    WVIM and LOCO default to a single refit per resampling iteration on the argument that budget is better spent on more iterations, and a batch_size argument dispatches multiple refits per benchmark call so they can run in parallel. Samplers gain samples_per_row for drawing multiple perturbations per observation.

    View source ↗
  2. 5mo ago

    Pre-trained learners supported; distribution-free inference added

    PFI, CFI, RFI and SAGE can now use an already-trained learner given a single-iteration resampling, removing the refit cost from workflows where the model already exists. A ci_method = "lei" option adds distribution-free inference on observation-wise loss differences.

    View source ↗
  3. 6mo ago

    Version bumped to mark the package as released

    Largely a signalling release: documentation cleanup, a SAGE min_permutations default raised to stop spurious early stopping, and removal of a comparison article now maintained elsewhere.

    View source ↗
  4. 9mo ago

    Confidence intervals arrive for feature importance scores

    The release that made uncertainty a first-class part of the output: ci_method offers raw, Nadeau-Bengio, quantile and CPI variance estimation, with standardisation and difference-versus-ratio options, and computation moves out of $compute() to avoid recomputing predictions.

    View source ↗