What Is the Fairness-Accuracy Trade-off?

Q: What Is the Fairness-Accuracy Trade-off?

The fairness-accuracy trade-off is the observation that fairness-mitigation strategies which reduce a model's fairness gap often also reduce its predictive accuracy, and the two cannot always be improved together - a model sits somewhere on a Pareto frontier between them, not at an independently-optimal point on each.

Closing a fairness gap is rarely free - the same intervention that shrinks a demographic parity gap can quietly cost the model a meaningful share of its accuracy, and neither number tells the whole story alone.

The One-Sentence Definition

The fairness-accuracy trade-off is the observation that fairness-mitigation strategies which reduce a model's fairness gap often also reduce its predictive accuracy, and the two cannot always be improved together - a model sits somewhere on a Pareto frontier between them, not at an independently-optimal point on each.

Why It Matters

Pre-, In-, and Post-Processing Fairness Mitigations walks through this repo's five-strategy ladder and what each stage does to the fairness gap. It never mentions accuracy at all. That is the other half of the same story: every one of those strategies also changes model performance, and treating "we closed the gap" as the end of the evaluation - without asking what it cost - is exactly how a fairness intervention can look like an unambiguous win when it was actually a trade a team never explicitly agreed to make.

This isn't a reason to avoid fairness mitigation. It's a reason to report both numbers side by side, every time, so the trade being made is visible and can be judged deliberately rather than discovered later.

Concrete Example: COMPAS - Audit 01

Both tables below use the baseline logistic regression model on race, across all five strategies, using the frozen numbers exactly as faircode/benchmark.py computed them (paper/results-frozen/results_fairness.csv and paper/results-frozen/results_performance.csv).

Strategy	Demographic Parity Diff	Accuracy	Accuracy vs. S0
S0 baseline	0.855 (sig.)	65.3%	-
S1 unawareness	0.172 (sig.)	57.1%	-8.2 pts
S2 unawareness_proxy_removal	0.115 (sig.)	57.2%	-8.1 pts
S3 in_processing	-0.014 (n.s.)	53.9%	-11.4 pts
S4 post_processing	0.023 (n.s.)	55.7%	-9.7 pts

The pattern Mitigation Strategies already shows on the fairness side - a huge S0→S1 improvement, a smaller S1→S2 gain, and S3/S4 converging near zero - has a cost sitting directly underneath it that a fairness-only table hides completely:

The single largest accuracy drop (S0 → S1, -8.2 points) coincides with the single largest fairness gain (85.5 → 17.2 points closed). Simply dropping the protected attribute buys most of the fairness improvement this ladder achieves, and most of the accuracy cost too, in the same step.

S3 in_processing reaches the best fairness number (-1.4 points, not statistically significant) at the worst accuracy (53.9%, an 11.4-point drop from baseline). This is the sharpest illustration of the trade-off on this audit: the strategy that gets closest to eliminating the demographic parity gap is also the one that costs the model the most predictive accuracy - not a coincidence, since ExponentiatedGradient is directly trading classification accuracy against the fairness constraint during training, by design.

S2 and S4 show it isn't a strict, monotonic trade. S2 has a smaller fairness gap than S1 (11.5 vs. 17.2 points) at essentially the same accuracy (57.2% vs. 57.1%) - a case where fairness improved with no meaningful accuracy cost. S4 recovers some accuracy back from S3 (55.7% vs. 53.9%) while its fairness gap stays statistically indistinguishable from S3's. The frontier isn't a straight line where every fairness point costs a fixed accuracy price; some steps are close to free, and some are expensive, which is exactly why both numbers need checking at every step rather than assuming the relationship is linear.

Detection Code

Pairs each strategy's fairness gap with its accuracy so a table can't report one without the other.

import pandas as pd


def fairness_accuracy_tradeoff(fairness_df, performance_df, audit, fairness_metric,
                                protected_attribute, model="logistic_regression"):
    """
    Joins one fairness metric's progression across the five strategies with
    the corresponding accuracy for the same audit/model, so a fairness gain
    can be checked against its accuracy cost at every step.

    Parameters:
        fairness_df: a DataFrame shaped like paper/results-frozen/results_fairness.csv
        performance_df: a DataFrame shaped like paper/results-frozen/results_performance.csv
        audit, fairness_metric, protected_attribute, model: filter values

    Returns a DataFrame with one row per strategy, in S0-S4 order, with the
    fairness value/significance and accuracy side by side, plus an
    `accuracy_delta_from_baseline` column.
    """
    order = ["baseline", "unawareness", "unawareness_proxy_removal",
             "in_processing", "post_processing"]

    fair = fairness_df[
        (fairness_df["audit"] == audit)
        & (fairness_df["metric"] == fairness_metric)
        & (fairness_df["protected_attribute"] == protected_attribute)
        & (fairness_df["model"] == model)
    ].set_index("strategy").reindex(order)[["value", "significant"]]
    fair.columns = ["fairness_value", "fairness_significant"]

    acc = performance_df[
        (performance_df["audit"] == audit)
        & (performance_df["metric"] == "accuracy")
        & (performance_df["model"] == model)
    ].set_index("strategy").reindex(order)[["value"]]
    acc.columns = ["accuracy"]

    joined = fair.join(acc).reset_index()
    joined["accuracy_delta_from_baseline"] = joined["accuracy"] - joined["accuracy"].iloc[0]
    return joined


# Usage example:
# import pandas as pd
# fairness = pd.read_csv("paper/results-frozen/results_fairness.csv")
# performance = pd.read_csv("paper/results-frozen/results_performance.csv")
# print(fairness_accuracy_tradeoff(
#     fairness, performance, "compas", "demographic_parity_diff", "race",
# ))

Limitations

1. The trade-off isn't fixed, universal, or always present

As S2 shows above, a fairness gain can arrive at essentially no accuracy cost. The trade-off is a common empirical pattern on this benchmark, not a law - never assume a fairness fix must cost accuracy without checking, and never assume a specific ratio (e.g. "1 fairness point costs 0.1 accuracy points") transfers from one audit or metric to another.

2. "Accuracy" is itself only one performance metric, and can hide its own asymmetry

A dropped-accuracy number doesn't say which predictions got worse. faircode/benchmark.py also computes AUC and F1 alongside accuracy for exactly this reason - a strategy that trades away F1 while accuracy looks stable is a different trade than one that drops all three metrics together. See What Is Accuracy Equality? for the parallel problem on the fairness side: a single blended number can hide two very different underlying stories.

3. The Pareto frontier is estimated from a specific dataset and split

The exact numbers above hold for the COMPAS audit's baseline logistic regression model and 80/20 stratified split (random_state=42) - a different model family, a different dataset, or a different split can land at a different point on the frontier. Treat the pattern (check both numbers, expect some cost, don't assume it's linear) as the transferable lesson, not the specific -8.2/-11.4/-9.7 point figures.

4. A statistically insignificant fairness gap is not the same as "fixed"

S3's -1.4-point gap and S4's 2.3-point gap both have confidence intervals crossing zero (see What Is a Bootstrap Confidence Interval?) - "not statistically significant on this test set" is weaker evidence than "the disparity is gone," and shouldn't be reported as though the accuracy cost bought a fully solved fairness problem.

What Are Pre-, In-, and Post-Processing Fairness Mitigations? - the fairness-only half of the exact same S0-S4 progression used here.
What Is Accuracy Equality? - a fairness metric built on accuracy itself, and the parallel reminder that one blended accuracy number can hide asymmetry underneath it.
What Is Bias-Variance Trade-off? - a different, model-internal trade-off; useful for not confusing "the model overfit" with "the model traded accuracy for fairness."

COMPAS/ - the audit above, where S3 buys the best fairness number at the worst accuracy.
faircode/benchmark.py - runs every strategy/model/metric combination in one uniform pipeline, producing both results_fairness.csv and results_performance.csv from the exact same train/test split so the two are directly comparable, strategy for strategy.