Out-of-fold explainability

Current mllabiome explainability is out-of-fold by default. This is the recommended mechanism for nested-CV and LODO experiments.

Principle

Every explained row is explained using a model that was not fitted on that row.

For MPMA-B, Baseline RF, and specific MPMA targets:


outer train → fit transformation → fit learner
outer test  → transform with the train-fitted transformation → explain

For MPMA-E:


for each outer fold:
  fit each selected member on outer train
  predict/explain outer test
  aggregate member predictions with the selected ensemble rule

Methods

By default, the configured target is explained using:

Method	OOF behavior
SHAP	Background from the outer-training fold; explained rows from the outer-test fold
LIME	Explainer fitted on outer-training transformed data; local explanations for outer-test rows
Permutation importance	Computed on the outer-test fold with `y_test`
ALE	Fold-fitted model and outer-test matrix; sparse features that are constant in a fold are skipped for that fold only
Interactions	Two-dimensional ALE interaction analysis, aggregated across folds

The package does not silently replace a requested method with another method. If a dependency is missing or the selected target cannot be explained, the explainability stage fails. Sparse ALE features are not a fallback case: they are recorded as not estimable for the affected fold and reported in diagnostics.

Target selection


EXPLAINABILITY = mll.Explainability(targets="auto", top_k=30)

targets="auto" runs:


MPMA-E, when selected_unit.json exists
MPMA-B
strict Baseline RF, when the required baseline exists

Specific target examples:


# Best individual MPMA only
EXPLAINABILITY = mll.Explainability(targets=("mpma_b",), top_k=30)
 
# Selected ensemble only
EXPLAINABILITY = mll.Explainability(targets=("mpma_e",), top_k=30)
 
# Strict baseline RF only
EXPLAINABILITY = mll.Explainability(targets=("baseline_rf",), top_k=30)
 
# Specific config ID
EXPLAINABILITY = mll.Explainability(targets=("48e21e491927",), top_k=30)

Important output files

File	Meaning
`feature_importance.tsv`	Consensus importance table across configured methods
`feature_importance_<method>.tsv`	Method-specific importance table
`oof_predictions.tsv`	Out-of-fold predictions used by explainability
`explained_unit.json`	Metadata for the explained target, including OOF status
`ale_skipped_features.tsv`	Sparse features/folds where ALE was not estimable, if any
`feature_interactions_by_target_all_methods.csv`	Interaction summary by target and method