[ChrisW09]

Fixes #83.

With unsupervised placement (target_aware=False, "uniform" or "quantile"), ReLUExpansionTransformer put its last center exactly at the training maximum. The ramp max(0, x - c) of that center was zero on every training row, so output_dim=k gave only k - 1 informative columns per feature. This happened on every fit of the standalone transformer, whose default placement is unsupervised "quantile", and of Preprocessor(numerical_method="relu", target_aware=False, ...).

Root cause

BaseCenterExpansion.fit placed the centers of every feature map through RBFPlacementAdapter.get_centers, which uses endpoint-inclusive locations (include_endpoints=True). That convention suits kernels that respond on both sides of a center (RBF, sigmoid, tanh). For the one-sided ReLU ramp, a center at the maximum is degenerate.

Fix

• BaseCenterExpansion.fit takes the unsupervised centers from a new hook, _unsupervised_centers(adapter, x, n_centers). The default implementation makes the same adapter.get_centers(x, None, n_centers, n_centers) call as before. The target-aware path still calls adapter.get_centers with the adaptive bounds.
• ReLUExpansionTransformer._unsupervised_centers places n_centers + 1 range-spanning locations and drops the right endpoint. The first center stays at the training minimum, which gives the linear term, and no ramp starts at the maximum.

Behaviour changes

• Unsupervised ReLU centers move. Uniform centers are linspace(min, max, output_dim + 1)[:-1], and quantile centers are the matching quantiles without the 100th percentile. Every column is now non-zero on the training data.
• The output width is unchanged. A zero-range feature still gets output_dim equal centers.
• RBF, sigmoid, tanh and the target-aware ReLU path are unchanged, and the golden regression baselines pass unchanged.

Documentation

• ReLUExpansionTransformer docstring: on the unsupervised path, the centers are the first output_dim of output_dim + 1 range-spanning locations, and why the right endpoint is dropped.
• docs/representations/functional_expansions.md, ReLU entry: the same in prose.
• Docstrings of the new hook BaseCenterExpansion._unsupervised_centers and of the ReLU override.

Tests

All tests are in tests/expansion/functional/test_reluexpansion_transformer.py.

• test_unsupervised_relu_has_no_dead_column[uniform, quantile]: all 6 columns are non-zero on the training data and have rank 6, the first center is at the minimum and the last is below the maximum. A zero-range feature keeps output_dim equal centers. Fails on main, passes here.
• test_unsupervised_relu_uniform_centers_drop_the_right_endpoint: the centers equal linspace(min, max, 7)[:-1]. Fails on main, passes here.
• test_preprocessor_relu_has_no_constant_column[uniform, quantile]: the issue's Preprocessor case with output_dim=7 gives no constant column and rank 7. Fails on main, passes here.

Reproduction

Output of the script from #83 on this branch:

centers: [3.000e-03 1.907e+00 3.884e+00 5.580e+00 7.034e+00 8.513e+00] | training max: 9.972
per-column max on the training data: [9.969 8.065 6.089 4.392 2.938 1.459]
Preprocessor relu: constant columns [] | rank 7

Checks

• just test: 1443 passed, 59 skipped, 7 xfailed; coverage 93.66% (gate 90%).
• Minimum dependencies (CI job: Python 3.10, numpy 1.24.4, pandas 2.0.3, scipy 1.10.1, scikit-learn 1.6.0): 1430 passed, 72 skipped, 7 xfailed; scripts/quickstart.py passes.
• just docs: strict build succeeds.
• just check: all pre-commit and pre-push hooks pass (ruff format, ruff lint, prettier, file hygiene, pyright with 0 errors).
• Import direction between subpackages is unchanged.

Merge notes

• Base: main.
• The PR for Unsupervised quantile placement repeats RBF/ReLU/sigmoid/tanh centers on tied data, duplicating columns #82 (fix(placement): return distinct quantile locations on tied data #119) also changes pretab/expansion/functional/relu.py, in the placement_strategy docstring. It merges cleanly with this branch. ReLU with "quantile" centers on tied data needs both fixes. On zero-inflated, discrete and top-coded features at output_dim=8:
  • With only this PR, tied quantiles still repeat. The top-coded feature, for example, keeps 2 zero columns and has 7 of 8 distinct columns.
  • With only Unsupervised quantile placement repeats RBF/ReLU/sigmoid/tanh centers on tied data, duplicating columns #82, the centers are distinct, but the last ramp still starts at the maximum, so one column is zero on the training data.
  • With both, no column is zero and all 8 are distinct.
• The PR for An invalid task such as "Regression" is silently treated as classification by the splines and selectors #93 (fix(validation): reject an unknown task instead of using classification #128) also changes pretab/expansion/functional/base.py and this test file, in other places, and merges cleanly.

This change was previously part of #86, which is closed in favour of one pull request per issue.