[Swapnil-byte-798]

What changed

For DropHighPSIFeatures(threshold="auto"), the chi-squared threshold used the requested number of bins even when equal-frequency discretization dropped duplicate quantile boundaries. Calculate the degrees of freedom from each feature's fitted bin edges.

Reproduction

With 100 rows in each split, change a binary-valued numerical score from 50% ones in the basis to 73% ones in the test split. Requests for two and ten bins both fit the same two bins and give PSI ≈ 0.229. The current threshold is ≈ 0.217 with bins=2 and ≈ 0.558 with bins=10, so only the former flags drift. Both requests now use ≈ 0.217.

The regression checks the numerical PSI, the feature selected for removal, and the transformed dataframe. The ten-bin case failed before the patch. All 177 tests in the focused PSI and discretization suites pass locally. Black and git diff --check pass.

The numerical bin-count calculation may also be relevant to the migration in #1089.

Codex assisted with reproduction, implementation, and tests.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.