Bug fixes in GCM

[bloebp]

Fixes silent wrong-result bugs (GMM sampling correlation, bootstrap CI percentiles, distribution-change early stopping, unit_change intercept, adaptive RBF kernel, discretization entropy, uniform noise std, categorical mode collapse), crashes and dtype issues (multiclass log_loss, max_num_samples subsampling, is_categorical with NaN, int32 overflow, empty-graph refutation, quantile guard, single-player Shapley, zero-root DAG, positional eq-parser args), a wrong class mapping (bayesiangaussianmixture, weighted clone), and a dead warning. Also raises a clear error for categorical counterfactual abduction instead of a leaked AttributeError. Adds regression tests for each.

[github-actions]

🤖 This is an automated response from Repo Assist.

Great work, @bloebp — this is a comprehensive collection of GCM correctness fixes with regression tests for each. The breadth of the changes (GMM sampling, bootstrap CI, distribution-change stopping, adaptive RBF, is_categorical NaN handling, single-player Shapley, categorical counterfactual abduction, and more) makes this a really valuable PR.

One heads-up for maintainers: my PR #1662 (fix(gcm): move NaN filter outside column loop in is_categorical()) addresses the same is_categorical NaN-indexing bug that is included here. Now that #1664 is open and covers the same issue (plus many more), PR #1662 should be closed in favour of this one. I've noted this in my memory and will close it once this PR merges (or maintainers can close #1662 directly).

A minor observation on the is_categorical patch in this PR: the current fix correctly hoists non_nan_X = X[~nan_mask] before the loop, which is exactly the right fix. The only difference from #1662 is that #1662 also restores the original loop variable name (X → non_nan_X throughout), which avoids shadowing the outer X. Both approaches are correct.

Looking forward to seeing this merged! 🎉

Generated by 🌈 Repo Assist, see workflow run. Learn more.

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/repo-assist.md@11c9a2c442e519ff2b427bf58679f5a525353f76

[Copilot]

Pull request overview

This PR addresses a broad set of correctness bugs in DoWhy’s GCM module (wrong results, crashes, and dtype edge cases) and adds targeted regression tests to prevent reintroductions.

Changes:

• Fix multiple silent wrong-result issues in sampling, bootstrap confidence intervals, entropy estimation, distribution-change scoring, and unit-change attributions.
• Harden several APIs against edge cases (empty-graph validation, single-player Shapley, categorical counterfactual abduction, NaNs in categorical detection, quantile input validation).
• Add/adjust regression tests covering each fixed scenario.

Reviewed changes

Copilot reviewed 33 out of 33 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/gcm/util/test_general.py	Adds regression test for informative error on categorical data with NaNs.
tests/gcm/test_whatif.py	Adds regression test for categorical-node counterfactual abduction error message.
tests/gcm/test_validation.py	Adds regression test for edgeless DAG refutation with FDR correction.
tests/gcm/test_unit_change.py	Adds regression test ensuring intercept differences are attributed in linear unit change.
tests/gcm/test_uncertainty.py	Updates entropy expectations and adds regression tests for small-range and constant data.
tests/gcm/test_stochastic_models.py	Adds regression test for spurious correlation avoidance in GMM sampling.
tests/gcm/test_stats.py	Adds regression test for quantile input validation error message.
tests/gcm/test_shapley.py	Adds regression test for single-player subset sampling Shapley.
tests/gcm/test_equation_parser.py	Adds tests for bayesiangaussianmixture mapping and positional arg parsing error.
tests/gcm/test_data_generator.py	Adds regression tests for zero-roots DAG generation and uniform noise std.
tests/gcm/test_confidence_intervals.py	Updates expected bootstrap CI bounds to correct two-sided percentiles.
tests/gcm/test_confidence_intervals_cms.py	Updates expected bootstrap CI bounds to correct two-sided percentiles.
tests/gcm/test_anomaly_scorers.py	Adds regression test for finite anomaly scores when MAD is zero.
tests/gcm/ml/test_classification.py	Adds regression test ensuring weighted classification clone preserves type/weights support.
dowhy/gcm/validation.py	Avoids FDR adjustment path on empty p-value lists.
dowhy/gcm/util/general.py	Fixes NaN-mask handling for multi-column categorical detection and raises clearer error.
dowhy/gcm/unit_change.py	Adds intercept support to linear mechanisms and attributes intercept deltas.
dowhy/gcm/uncertainty.py	Fixes discretization binning for small ranges; handles constant-data entropy.
dowhy/gcm/stochastic_models.py	Shuffles GMM samples to prevent spurious correlations between independently sampled nodes.
dowhy/gcm/stats.py	Tightens quantile validation to enforce (0, 1].
dowhy/gcm/shapley.py	Fixes subset-sampling approximation for single-player case.
dowhy/gcm/model_evaluation.py	Fixes max_num_samples subsampling and categorical mode shape handling.
dowhy/gcm/ml/classification.py	Ensures weighted classifier clone returns weighted clone.
dowhy/gcm/influence.py	Fixes infinity check for results stored as Python list.
dowhy/gcm/independence_test/kernel_operation.py	Fixes per-column adaptive RBF distance computation; updates delta-kernel handling.
dowhy/gcm/equation_parser.py	Fixes bayesiangaussianmixture mapping; improves argument parsing and parent-name collision warning.
dowhy/gcm/distribution_change.py	Fixes early stopping logic to track convergence of running average.
dowhy/gcm/data_generator.py	Fixes zero-root DAG generation and corrects uniform noise std parameterization.
dowhy/gcm/confidence_intervals.py	Fixes percentile bootstrap CI bounds to use two-sided percentiles.
dowhy/gcm/causal_mechanisms.py	Prevents int32 overflow by using int64 in discrete ANM.
dowhy/gcm/auto.py	Fixes multiclass log_loss usage by passing explicit label set.
dowhy/gcm/anomaly_scorers.py	Avoids division-by-zero in MAD-based scorer by falling back to EPS.
dowhy/gcm/_noise.py	Raises clear error when noise cannot be recovered for non-invertible mechanisms.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[bloebp]

@emrekiciman any concerns left?