Tree refactor

[Advueu963]

Motivation and Context

The shapiq.tree module had grown organically into a single-file/flat layout that was increasingly hard to extend. This PR does three things simultaneously:

1. Refactors shapiq.tree into a clean subpackage — separating conversion logic, explanation algorithms, validation, and the new algorithms into dedicated submodules. This makes the codebase easier to maintain, test, and
   extend.
2. Adds ProxySHAP — a new approximator that uses a lightweight proxy tree model (XGBoost by default) to compute interactions cheaply and then corrects for the approximation error on the true model. This is particularly
   useful when the true model is expensive to query.
3. Adds LinearTreeSHAP — an efficient first-order-only Shapley value implementation based on the Yu et al. (2022) NeurIPS paper, backed by a dedicated C++ extension for high throughput.

Along the way, InterventionalGame gains multi-class LightGBM support (a correctness fix), IsolationForest is added to the list of supported sklearn tree models, and InterventionalTreeExplainer is ported to use batched
sparse C-extensions for faster tree traversal.

---

Public API Changes

• No Public API changes
• Yes, Public API changes (Details below)

New exports:

• shapiq.approximator.ProxySHAP — proxy-model-accelerated interaction estimator
• shapiq.approximator.MSRBiased — biased multilinear-extension sampling regression (also used internally by ProxySHAP)
• shapiq.tree.LinearTreeSHAP — fast first-order Shapley values for tree models (index="SV" only)
• shapiq.tree.InterventionalTreeExplainer — previously internal, now a first-class export
• shapiq.tree.InterventionalGame — previously internal, now a first-class export

Unchanged interfaces:

• shapiq.TreeExplainer — public interface is fully preserved; the refactor is internal only.

New model support:

• sklearn.ensemble.IsolationForest is now accepted by TreeExplainer and validate_tree_model.

---

How Has This Been Tested?

• Unit tests cover ProxySHAP, MSRBiased, and LinearTreeSHAP for correctness of outputs and supported index types.
• The InterventionalGame LightGBM multi-class fix is tested with a multi-class LGBMClassifier.
• The refactored shapiq.tree subpackage is tested end-to-end via the existing TreeExplainer test suite, which exercises all supported model types (sklearn trees, XGBoost, LightGBM) and interaction indices.
• IsolationForest support is covered by a dedicated test.
• All tests run locally via uv run pytest tests/shapiq.

---

Checklist

• The changes have been tested locally.
• Documentation has been updated (if the public API or usage changes).
• An entry has been added to CHANGELOG.md (if relevant for users).
• The code follows the project's style guidelines.
• I have considered the impact of these changes on the public API.

---

[mmschlk]

Many changes which we tested quite strongly. I am still expecting some issues we need to find before release.

[codecov]

Codecov Report

❌ Patch coverage is 87.57485% with 83 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/shapiq/tree/interventional/explainer.py	84.02%	31 Missing ⚠️
src/shapiq/approximator/proxy/proxyshap.py	86.81%	12 Missing ⚠️
src/shapiq/tree/base.py	80.64%	12 Missing ⚠️
src/shapiq/tree/conversion/boosting.py	74.28%	9 Missing ⚠️
src/shapiq/tree/conversion_old/lightgbm.py	0.00%	5 Missing ⚠️
src/shapiq/tree/interventional/game.py	92.68%	3 Missing ⚠️
src/shapiq/tree/linear/explainer.py	96.34%	3 Missing ⚠️
src/shapiq/tree/__init__.py	81.81%	2 Missing ⚠️
src/shapiq/tree/validation.py	90.47%	2 Missing ⚠️
src/shapiq/tree/conversion/__init__.py	96.15%	1 Missing ⚠️
... and 3 more

📢 Thoughts on this report? Let us know!