Fix Greenacre correction for subset MCA (#206)#236
Open
MaxHalford wants to merge 8 commits into
Open
Conversation
When `one_hot_columns_to_drop` is set, the closed-form Benzécri/Greenacre formula no longer applies — it assumes uniform row sums in the indicator matrix, which subsetting violates. Switch to Greenacre's subset-MCA adjustment (CA in Practice, ch. 21), ported from R's `ca::mjca(lambda='adjusted', subsetcat=...)`. The fix only triggers when `one_hot_columns_to_drop is not None` and `correction='greenacre'`; the existing non-subset path is unchanged. Fixes #206. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Build the indicator matrix directly as a scipy CSC sparse matrix (factorize columns, build COO from offsets) instead of going through pd.get_dummies. Densify into one contiguous float ndarray right before CA.fit so sklearn's check_array sees a single block instead of iterating over J per-column ExtensionArrays. The subset-Greenacre Burt matmul now also benefits from sparse Zᵀ Z. Benchmarks (n_components=5, sklearn engine, 3 runs / best): 1k x 10 x 5 : 23ms -> 4ms (-83%) 10k x 20 x 10 : 62ms -> 45ms (-27%) 50k x 30 x 10 : 413ms -> 319ms (-23%) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Resolves ty unresolved-attribute warnings for ``full_one_hot_sp.T`` and ``np.repeat(..., n_levels)``: both are only used when ``self.one_hot`` is true, so move the block inside that branch instead of carrying ``None`` sentinels. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
FactoMineR has no eigenvalue correction; the existing test_abdi_2007_correction relies on hardcoded numbers from the paper. Add a live rpy2 cross-check against ca::mjca(lambda='adjusted'), which implements the same closed-form correction prince does for the non-subset path. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
scipy.linalg.svd (and fbpca) cap at min(M, N) singular values, so for a matrix of rank r < n_components the existing slice ``s = s[:n_components]`` silently returned fewer than the requested number of components, causing downstream shape mismatches (e.g. MFA with n_components=3 on a 2-column numerical group). Normalise all engines to return exactly n_components by padding U/s/V with zeros for the missing tail (sklearn's randomized_svd already does this implicitly with noisy padding). The extra components have zero singular values, so their contribution to coordinates and inertias is zero — matching the rank-deficient mathematical answer. With the scipy engine no longer crashing, switch the TestMFACategorical fixture to engine="scipy" so tight (atol=1e-4) comparisons against FactoMineR don't flake on randomized-SVD precision noise on the last component (CI flake on PR #236). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
one_hot_columns_to_dropis used withcorrection="greenacre", the existing closed-form correction is mis-calibrated: it assumes uniform row sums in the indicator matrix, which dropping categories breaks. Reported in Correction for the explained inertia in subset MCA #206.ca::mjca(lambda='adjusted', subsetcat=...): column marginals are computed on the full Burt matrix (so dropping categories doesn't redistribute mass), and per-dimension inertia is taken from the eigendecomposition ofS_nullrestricted to the active categories, with the total inertia coming fromS_e(independence within each variable's block) on the same subset.one_hot_columns_to_drop is not None+correction="greenacre". The existing non-subset Greenacre and Benzécri paths are unchanged.Test plan
test_subset_greenacre_matches_ca_mjcacomparesprince.MCAagainst Rca::mjcaon the burgundy wines dataset — eigenvalues and percentages match to ~1e-8.test_abdi_2007_correction(non-subset Greenacre/Benzécri doctest) still passes.tests/test_mca.pypasses (60/60).🤖 Generated with Claude Code