---

name: interpretable-models description: "Choose and operate AIX360's directly interpretable models, rule learners, prototypes, model differencing, teaching explanations, and optional neural interpretable methods." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Interpretable models

Use this skill when the request is for a model whose prediction mechanism is itself inspectable, a compact prototype set, an explicit rule list, a comparison of two models, a label-plus-teaching explanation, or a conceptual neural interpretability method. This skill covers AIX360 0.3.0 APIs and their optional dependency boundaries.

Route before coding

• Explain an already trained black box with LIME, SHAP, GroupedCE, or a neural contrastive method: route to ../local-black-box/SKILL.md.
• CEM, recourse, certification, or optimal transport: route to ../counterfactual-and-certification/SKILL.md.
• Time-series explainers: route to ../time-series/SKILL.md.
• Dataset loading, preprocessing, or metric contracts: route to ../datasets-and-metrics/SKILL.md.

Do not call a rule learner an explainer of an existing model: BRCG, GLRM, RIPPER, CoFrNet, and TED fit a predictive mechanism from labeled data. IMD consumes outputs from two already-fitted models. ProtoDash selects exemplars from a candidate set and can summarize data or support a local comparison. DIPVAE edits a learned latent generative model. ProfWeight trains a simpler model using confidence-derived sample weights.

Operating sequence

1. State the object being produced: prototypes, binary/regression rule model, DNF rules, model-difference regions, label/explanation pair, or a latent or confidence-profile model.
2. Freeze the training schema. Keep column names and ordering stable; retain the fitted FeatureBinarizer/FeatureBinarizerFromTrees for every later transform. Decide how missing values are represented before fitting.
3. Select the smallest dependency family in api-reference.md. Treat optional legacy families as unavailable unless their imports and versions are proven.
4. Fit only on training data, then predict or explain on a separately transformed/evaluated set. Check output shape, labels/classes, feature names, and rule count before interpreting the result.
5. Save a portable semantic summary (rule strings, prototype source indices and weights, diff regions, or latent edit settings), not an opaque live object.
6. Report accuracy/coverage or other metrics separately from readability. A compact model is not automatically faithful to another model.

Fast choices

• Prototype selection: ProtodashExplainer.explain(X, Y, m, ...) returns (weights, indices, objective_history); Y is the candidate pool and X is the data being summarized/explained.
• Binary rule classifier: binarize a pandas frame, fit BooleanRuleCG through BRCGExplainer, then call predict and explain.
• Interpretable regression/classification: use LinearRuleRegression or LogisticRuleRegression through GLRMExplainer; call predict (and predict_proba for logistic) and inspect the coefficient DataFrame.
• Native rule induction: fit RipperExplainer on a pandas DataFrame and Series, with target_label for a binary positive class; call explain() for a TRXF DnfRuleSet.
• Model differencing: obtain aligned predictions from two classifiers on the same pandas frame, fit IMDExplainer.fit(X, y1, y2, max_depth=...), and inspect diffrules, diffregions, and metrics.
• Teaching explanations: fit TED_CartesianExplainer(base_estimator) on (X, Y, E) where explanation ids are dense non-negative integers; use predict_explain, predict, and explain.
• Optional neural methods: use CoFrNet for continued-fraction network structure, DIPVAE for latent edits, and ProfWeight for confidence-profile weighted simple models only after their backend environments are validated.

Hard safety checks

• FeatureBinarizerFromTrees.fit requires y, rejects NaN/None during fit, and its transform schema is learned from selected input features. A missing selected column is a schema error, not a signal to silently impute.
• BRCG solves LPs and can emit no useful clauses; inspect w, z, and the rules list rather than assuming a non-empty explanation.
• ProtoDash weights are non-negative but unnormalized; zero, repeated, or numerically unstable weights need a smaller candidate set, finite scaled inputs, or the alternate optimizer.
• TRXF predicates use Feature, Predicate, Conjunction, and DnfRuleSet. Preserve typed values and operators when serializing; do not use repr as a persistence format.
• Do not promise compatibility for legacy TensorFlow/Keras 1.x ProfWeight or the optional PyTorch CoFrNet/DIPVAE code from a modern environment without a separate probe. See troubleshooting.md.

For concrete call signatures, output contracts, dependency boundaries, and recovery procedures, read the four bundled references. The tiny smoke helper is at scripts/feature_binarizer_smoke.py and is safe to run locally with --help before running it.