Skip to content

[T-072] Importance: HistGradientBoosting + SHAP (global, interactions, per-mode) + GP lengthscale cross-check #131

Closed

Description

@CodewithJha

Task T-072 · Milestone M6 — Analysis + risk · Size M · Priority P0 · Epic #6

Implements FR-ANA-003 · PR-06

Branch feat/t-072-importance-histgradientboosting-shap-global-interactions — one task = one branch = one PR (doc 32 §6)

Summary

Implement analysis/importance.py ImportanceExplainer.explain() per doc 09 §4.1–4.2: balanced HistGB classifier on u, 5-fold stratified CV (ROC-AUC, balanced accuracy, low-confidence flag if n_fail < min_failures_confident), mean |SHAP|, pairwise interaction strength via shap_interaction_values, ARD 1/ℓ_j comparison with disagreement flag, per-mode importance.

Context / motivation

Answers why failures happen and which parameter combinations matter (H3).

Scope

In scope

  • Fallback to GradientBoostingClassifier if SHAP/HistGB versions mismatch.
  • ImportanceReport model.
  • ImportanceSettings model with the new keys.

Implementation guidance

  • Module roboprobe/analysis/importance.py: class ImportanceExplainer: def __init__(self, settings: ImportanceSettings, seed: int) -> None; def explain(self, X: NDArray, y_fail: NDArray[np.bool_], names: list[str], gp: RobustnessGP | None = None, modes: Sequence[FailureModeSummary] = ()) -> ImportanceReport (doc 05 §7).
  • Model: HistGradientBoostingClassifier(max_depth, learning_rate, max_iter from analysis.gbm, class_weight="balanced", random_state=seed) on unit-space X of all valid runs (FAIL vs PASS).
  • Quality: StratifiedKFold(analysis.cv_folds, shuffle=True, random_state=seed) → ROC-AUC and balanced accuracy (mean ± std); low_confidence = n_fail < analysis.min_failures_confident or auc < 0.7 (AUC floor as setting analysis.min_cv_auc, new).
  • SHAP: shap.TreeExplainer(model); global importance = mean |SHAP|; interactions via shap_interaction_values on a sample of analysis.shap_max_samples (new, default 1000) rows → symmetric d×d strength matrix (off-diagonal mean |value|).
  • GP cross-check: 1/ℓ_j from the ARD kernel (T-055a) normalised to sum 1; flag parameters whose rank differs by ≥ 2 positions (disagreement=True).
  • Per-mode importance: one-vs-rest model per mode with ≥ min_failures_confident members.
  • ImportanceReport serialisable to JSON (lists, not arrays) for report.json and the /importance endpoint.
  • Fallback if SHAP's tree support doesn't cover the installed sklearn version: GradientBoostingClassifier + same explainer; record model_kind in the report.
  • Standards: respect the layering contracts (doc 16 §2) and conventions in doc 34; no hard-coded tunables — thresholds/defaults come from typed settings (doc 27); no magic numbers (ruff PLR2004); public APIs typed (mypy strict) and documented.

Configuration keys

From the configuration reference (doc 27). Keys marked (new) must be added to the typed settings models and appear in the generated reference (T-003c).

  • analysis.gbm
  • analysis.cv_folds
  • analysis.min_failures_confident
  • analysis.min_cv_auc (new)
  • analysis.shap_max_samples (new)

Acceptance criteria

  • Planted-parameter synthetic data: planted params ranked top-k; planted interaction pair has highest interaction strength.
  • Low-confidence flag works.
  • Report JSON validates against schemas/analysis_report.schema.json importance block (T-075a).
  • Low-confidence flag tested for both triggers.
  • Deterministic for fixed seed (two runs → equal JSON).
  • Runtime < 30 s for 2 000 runs, d = 8 (benchmark marker).
  • Tests listed below added and green in CI; coverage of new code meets the gate in doc 18 §3.
  • Docs, config reference and CHANGELOG Unreleased updated where behaviour, keys or interfaces changed.
  • Definition of Done (doc 32 §3) satisfied.

Tests required

  • tests/unit/analysis/test_importance.py.
  • Synthetic planted-parameter dataset generator in tests/fixtures/analysis/synthetic.py (shared with T-073).

Dependencies

Blocked by: #110 (T-055a)

Blocks: #132 (T-073), #137 (T-075a)

Parallelization notes

Files / paths owned by this task — avoid editing other paths; shared files (pyproject.toml entry points, cli/main.py, api/routers/__init__.py, router.tsx, docker/compose.yaml, ci.yml) are append-only:

  • backend/src/roboprobe/analysis/importance.py

Independent of clustering stream.

Definition of Done

Per doc 32 §3 and AGENTS.md §5: lint, format, mypy strict and import-linter green; tests added and passing; docs updated; PR template completed with Closes #<this issue>; Conventional Commit messages referencing T-072 (doc 16 §6); commits authored only by the repository owner identity (AGENTS.md §4); CI green; reviewer checklist (doc 32 §5.1) satisfied.

References

Design docs: doc 09 §4 · doc 09 §4.1 · doc 09 §4.2 · doc 18 §4.6 · ADR-0012 · doc 05 §7 · doc 27 §4.8

Requirements: FR-ANA-003 · PR-06 (traceability: doc 28)

External:

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:analysisFailure clustering, SHAP, rules, interactionsmilestone:M6M6 — Analysis + riskpriority:P0Critical path / blocks otherssize:M1-3 daystaskChild task of a milestone epictype:featureNew functionality

    Projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions