fix(regression): MAPE near_zero_values counts rows, not DataFrame cells

[feiiiiii5]

Problem

The MAPE metric tells the user how many rows were affected by near-zero targets, and that number is always twice the real count.

User-visible symptom (insight rendered next to the MAPE percentage-error plot, src/evidently/metrics/regression.py:356-366):

Near-zero values detected (|target| ≤ epsilon). Applied zero_handling='replace'. Affected rows — current: 6, reference: 4.

for a dataset that actually has 3 near-zero rows in current and 2 in reference.

The same doubled number is persisted in report snapshots as MoreRegressionMetrics.near_zero_values, so it is also part of the serialized API surface, not just the HTML text.

Root cause

src/evidently/legacy/calculations/regression_performance.py, _calculate_quality_metrics:

data = dataset[[prediction_column, target_column]]      # 2 columns
...
epsilon_values = data[~(abs(data[target_column]) > epsilon)]
...
"near_zero_values": epsilon_values.size,

epsilon_values is a DataFrame with two columns, and pandas.DataFrame.size is the number of cells (len(df) * len(df.columns)), not the number of rows. The filter always selects exactly the [prediction, target] pair, so .size is 2 * <affected rows>.

The counter was introduced in #1835 ("Add methods to handle near-zero values for MAPE metric"). RegressionPerformanceMetrics.near_zero_values is annotated int and every consumer treats it as a row count:

• src/evidently/metrics/regression.py:351 — if cur_near_zero_values + ref_near_zero_values > 0
• src/evidently/metrics/regression.py:362 — f" Affected rows — current: {cur_near_zero_values}"

Nothing depends on the doubled value, and no existing test asserted it.

Fix

Report rows instead of cells:

"near_zero_values": len(epsilon_values),

The other use of .size in the same function (epsilon_values.size > 0, a pure emptiness check) is left untouched — it is correct either way and changing it would be churn.

Reproduction / verification

Environment: macOS, Python 3.11.15, pytest 7.4.4, pandas 3.0.5, numpy 2.4.6, evidently installed editable from this branch.

Minimal trigger (calculate_regression_performance with 3 rows whose target is exactly 0.0):

expected near-zero rows: 3
reported near_zero_values: 6     # before the fix
reported near_zero_values: 3     # after the fix

New regression test tests/calculations/test_regression_performance.py.

Base commit 34771fe41c08ca074d96a52aaf4a45cd640cbaf5, test file added, fix not applied:

python -m pytest tests/calculations/test_regression_performance.py -v
...
E       AssertionError: assert 6 == 3
5 failed, 1 passed

Same command with the fix applied:

python -m pytest tests/calculations/test_regression_performance.py -v
...
6 passed

Control runs with an identical command, to separate pre-existing failures:

python -m pytest tests/calculations tests/tests/test_regression_performance_tests.py tests/utils tests/report -q

tree	command	result
base commit (new test not present)	the 4-path control command	178 passed
base commit + new test (fix line mutated back)	the 4-path control command	5 failed, 179 passed
this branch (fix + new test)	the 4-path control command	184 passed

The mutation run above is the base-code control for the new tests: on this branch I reverted only the fixed line back to epsilon_values.size and re-ran the same command — exactly the 5 count-sensitive tests fail and nothing else.

ruff format --diff reports both changed files as already formatted.

Scope and limitations

• Single-line behaviour fix plus a new test module. No API, schema or serialization format change; only the value written to the existing near_zero_values field becomes correct.
• Snapshots produced by earlier releases keep the doubled value they were written with; this PR does not rewrite existing snapshots.
• Local run used PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 because an unrelated third-party pytest plugin in the shared environment failed to import. The full suite was not run locally (it needs the bikes/scipy datasets and the [llm,spark] extras that CI downloads), so the control runs above are limited to tests/calculations, tests/tests/test_regression_performance_tests.py, tests/utils and tests/report.
• My local pandas/numpy are newer than the requirements.min.txt pins (pandas 3.0.5 / numpy 2.4.6 vs 1.3.5 / 1.23.0). The fix is a len() vs .size distinction, which is stable across those versions, but the matrix runs in CI are the authoritative check.
• Separately noticed while reading this code, not changed here: the MAPE.zero_handling docstring in src/evidently/metrics/regression.py says rows where "the target or prediction" is within ±epsilon are handled, while the implementation and the insight text only consider the target. That is an independent documentation/behaviour question and is left to the maintainers.