Explainability

The standard explainability workflow is out-of-fold: an observation is explained using a model that was not fitted on that observation.

Configure methods


EXPLAINABILITY = mll.Explainability(
    targets=("mpma_b", "mpma_e"),
    profile="standard",
    methods=(
        mll.Permutation(),
        mll.SHAP(),
        mll.ALE(),
        mll.LIME(),
        mll.ALEInteractions(),
    ),
    classes="auto",
    top_k=30,
)

targets="auto" can be used when the standard selected targets should be determined from available run artefacts.

OOF principle

For an individual MPMA:


outer train → fit transformation → fit learner
outer test  → transform using training-fitted state → explain

For MPMA-E, the selected member models are fitted on the outer training partition and their predictions or explanations are aggregated for the outer test observations.

Methods

Method	Interpretation
Permutation importance	Change in held-out performance when a feature is disrupted
SHAP	Model-specific additive attribution under the selected explainer
ALE	Average local response over observed feature ranges
LIME	Local surrogate explanation around selected observations
ALE interactions	Pairwise interaction structure estimated from the fitted model

Local explanation behavior can be configured with mll.LocalExplanations.

Scientific interpretation

Feature importance is evidence about model behavior, not proof that a taxon or variable is causal. Microbiome features are compositional, sparse, and correlated; attribution can therefore be redistributed across related features.

For biological claims, prioritize stability across outer folds and cohorts, agreement across complementary methods, and independent replication.