fairness_check() reports "No bias was detected!" when a subgroup ratio is exactly 0 or infinite (ratio masked to NaN before the epsilon comparison) #585

python/dalex/dalex/fairness/_group_fairness/utils.py (checked at commit 77ea3ad)

calculate_ratio() converts any subgroup/privileged metric ratio that comes out as exactly 0, or as inf (privileged metric is 0), into NaN:

def calculate_ratio(sub_confusion_matrix_metrics, privileged):
    """
    Calculates ratio of metrics - divides all by privileged
    Does not allow for zeros in ratios, instead puts NaN
    """
    ...
    df_ratio[df_ratio == 0] = np.nan          # utils.py:151
    df_ratio[np.isinf(df_ratio)] = np.nan     # utils.py:152
    ...

GroupFairnessClassification stores this frame as self.result (object.py:27,34). universal_fairness_check() then feeds it straight into the pass/fail comparison. There is no NaN-aware branch:

metrics_exceeded = ((metric_ratios > 1 / epsilon) | (epsilon > metric_ratios)).apply(sum, 0)   # utils.py:358
...
else:
    print("No bias was detected!")                                                              # utils.py:366

In numpy/pandas, both NaN > x and x > NaN evaluate to False. So a ratio of exactly 0 contributes 0 to metrics_exceeded and is never added to names_of_exceeded_metrics, even though it is the most extreme disparity a metric can show. An example is PPV of 0.8 for the privileged subgroup and 0.0 for the unprivileged one. An infinite ratio behaves the same way. The headline verdict ("is fair" / "No bias was detected!") depends only on that count, so both ratios are treated as within the (epsilon, 1/epsilon) band. The fairness_check() docstring defines a ratio outside that band as a failure.

The problem is wider than PPV. SubgroupConfusionMatrixMetrics rounds every metric to 3 decimals (utils.py:91), so any subgroup rate below 0.0005 becomes exactly 0.0 before the division and is masked as well. The most realistic trigger is a model that never predicts the positive class for one subgroup (tp = fp = 0). TPR, FPR and STP are then 0 for that subgroup (ratios masked to NaN), and PPV is NaN (0/0). Only the ACC ratio remains, and the verdict is "fair".

With verbose=True (the default), a NaN warning is printed after the verdict, together with the ratio table showing NaN. The Conclusion: line still says the model is fair, and with verbose=False nothing flags it.

Measured

The code was reproduced standalone: calculate_ratio(), calculate_parity_loss(), fairness_check_metrics() and universal_fairness_check() were copied verbatim into a bare Python process (numpy/pandas only, no dalex import). They were run in the same order GroupFairnessClassification uses (calculate_ratio() → universal_fairness_check()) with epsilon=0.8, the library default. All other metrics were identical between the two subgroups unless stated.

The unprivileged subgroup with the worse outcome (exactly 0) is the only PPV case that is not flagged. Every less extreme rate tested, from 0.001 to 0.75, is handled correctly by the same functions. test_calculate_ratio in test/test_fairness.py pins the zero→NaN conversion, which exists to keep np.log finite in calculate_parity_loss(). No test checks the fairness_check() verdict for a subgroup whose ratio is 0 or inf.

Consequence

GroupFairnessClassification.fairness_check() is the library's one-call fairness audit. Consider a subgroup whose confusion-matrix metric is exactly zero while the privileged subgroup's rate is nonzero, or the reverse. That subgroup is dropped from the "exceeded" tally and cannot push the verdict to "not fair" or "cannot be called fair". A subgroup 1000x less extreme (rate 0.001 instead of 0.000) is caught. A model that assigns no positive predictions at all to a protected subgroup can be reported as fair on statistical parity, equal opportunity and predictive equality.

The same masking affects calculate_parity_loss(). Its .apply(sum) uses Python's builtin, non-skipna sum, so one masked subgroup turns the whole metric's parity loss into NaN, even when other subgroups have real, computable disparities. Measured with three subgroups (PPV ratios 0 → NaN and 0.5): parity loss for PPV = nan.

Suggested fix

The codebase already has a pattern for surfacing NaN explicitly. _unwrap_parity_loss_data() in the same file checks any(pd.isna(data.score)) and reports which models are affected. PR #403 (fixing #399, "NA in metric_scores plot causes random lines to appear") added a similar explicit NaN branch to the fairness plotting code.

universal_fairness_check() should handle the masked ratios before the threshold comparison. A ratio that was 0 or inf before masking should either count as "exceeded" for that metric/subgroup, since it is further outside the epsilon band than any finite ratio, or produce an explicit "cannot be decided, review required" verdict. It should not silently count as "not exceeded". The masking in calculate_ratio() can stay as it is for the parity-loss log. The check only needs to tell apart a NaN from a 0/inf ratio and a NaN from an undefined metric (0/0), or look at the unmasked ratio.

Happy to open the PR.