Fairness across users, bodies, and demographics

"My aggregate accuracy was ninety-four percent, which everyone celebrated. Nobody asked which six percent, so I quietly kept it all in one corner of the population."

A Diplomatically Averaged AI Agent

Where the gaps come from

Unfairness in human activity recognition is rarely malice and almost always physics plus sampling. Three mechanisms dominate. First, dataset composition: public benchmarks and internal collections skew toward convenient subjects, typically younger, fitter, and disproportionately male, so the model simply sees fewer examples of everyone else and fits them worse. Second, physiological signal differences: the same activity produces genuinely different inertial signatures across bodies. Stride length scales with leg length, so a person of height \(1.55\ \mathrm{m}\) and one of \(1.90\ \mathrm{m}\) walking at the same speed generate different step cadences and peak accelerations; a threshold or feature tuned to the mean stride under-counts the short-legged walker. Cadence, ground-contact time, and vertical impact all shift with age, weight, footwear, and gait disorders such as Parkinsonian shuffling or a post-stroke hemiparetic swing. Third, sensor coupling: a wrist device sits and moves differently on a thin wrist than a large one, and a hip clip rides at a different angle on different body shapes, so even identical motion reaches the accelerometer transformed.

A fourth mechanism is modality-specific and easy to forget in an inertial chapter: any activity pipeline that also ingests optical heart rate inherits the skin-tone bias of photoplethysmography. Green light is absorbed more by higher-melanin skin, lowering the signal-to-noise ratio of the pulse and degrading any activity or energy-expenditure estimate that leans on heart rate, an effect examined in depth in Chapter 30. The point is that a single fairness gap can be caused by data, by biomechanics, by mounting, and by sensor physics at once, so there is no single lever that fixes all of it.

Measuring fairness without hiding the tail

Start by defining the groups explicitly. Let \(g \in \mathcal{G}\) index a protected or body-related attribute (sex, age band, height quartile, BMI band, reported mobility aid). For a per-subject or per-window accuracy \(\mathrm{acc}(g)\), two summaries matter more than the pooled mean. The gap is \(\Delta = \max_g \mathrm{acc}(g) - \min_g \mathrm{acc}(g)\), and the worst-group accuracy is \(\min_g \mathrm{acc}(g)\). For classification you usually want error-rate parity rather than raw accuracy parity, because base rates of activities differ across groups; report per-group false-negative rate on the safety-critical class (a missed fall, a missed seizure-like event) since a group with more misses is the group being failed. Crucially, none of this works unless the test set is split by subject and stratified so every group is actually represented on the held-out side, the leakage discipline of Chapter 5 applied group by group.

The listing below computes per-group accuracy, the disparity gap, and the worst-group number from arrays of true labels, predictions, and a group tag per sample. It is deliberately small because the audit itself should be trivial to run on every model version.

import numpy as np

def fairness_report(y_true, y_pred, groups):
    y_true, y_pred, groups = map(np.asarray, (y_true, y_pred, groups))
    per_group = {}
    for g in np.unique(groups):
        m = groups == g
        per_group[g] = (y_true[m] == y_pred[m]).mean()
    overall = (y_true == y_pred).mean()
    worst = min(per_group.values())
    gap = max(per_group.values()) - worst
    return overall, worst, gap, per_group

# Toy data: majority group "M" is easy, minority "F_short" is under-served.
rng = np.random.default_rng(0)
groups = np.array(["M"] * 800 + ["F_short"] * 200)
y_true = rng.integers(0, 5, size=1000)
y_pred = y_true.copy()
flip_M = (groups == "M") & (rng.random(1000) < 0.05)      # 5% error
flip_F = (groups == "F_short") & (rng.random(1000) < 0.30) # 30% error
y_pred[flip_M | flip_F] = rng.integers(0, 5, size=(flip_M | flip_F).sum())

overall, worst, gap, pg = fairness_report(y_true, y_pred, groups)
print(f"overall={overall:.3f}  worst-group={worst:.3f}  gap={gap:.3f}")
print({k: round(v, 3) for k, v in pg.items()})

As Listing 26.6 makes plain, the pooled average and the worst-group number tell different stories about the same model, and only the second one describes the user you are most likely to harm. The same skeleton extends to false-negative rate per group by swapping the accuracy line for a per-class miss rate.

Closing the gap

Once you can see a gap, several levers narrow it, roughly in order of cost. The cheapest is evaluation discipline: making worst-group score a first-class release gate so a regression that helps the majority but hurts a minority cannot ship unnoticed. Next is data: targeted collection to raise the density of under-served bodies, and group-balanced or reweighted sampling so training loss stops being dominated by the majority. A principled version of reweighting is group distributionally robust optimization (Group DRO), which minimizes the loss of the worst group rather than the average, optimizing \(\min_\theta \max_{g} \mathbb{E}_{(x,y)\sim g}\,[\ell_\theta(x,y)]\); it directly targets \(\min_g \mathrm{acc}(g)\) instead of the mean and is a small change to the training loop. A complementary, powerful lever is personalization: the adaptation methods of Section 26.4 bend a general model toward the individual wearing it, which dissolves many between-group gaps because a model adapted to one short-stride older user no longer needs the population to have represented her. Personalization is often the most effective fairness intervention available, since the ultimate minority group is a single person.

Two cautions keep this from becoming naive. First, mitigation can trade average for worst-group; that trade is usually the right one for a safety product, but state it explicitly rather than reporting only the number that improved. Second, fairness fixes must survive the same leakage scrutiny as everything else, and they interact with governance: many jurisdictions and standards now expect documented subgroup performance for health-adjacent wearables, the responsible-deployment terrain of Chapter 70.

Auditing without fooling yourself

Two failure modes turn a fairness audit into theater. The first is underpowered groups: a per-group accuracy computed on eight subjects has a confidence interval so wide that the reported gap may be pure sampling noise. Report the number of subjects per group and a confidence interval; if a group is too small to measure, that is itself a finding that you have not collected enough of that population to make any claim. The second is intersectionality: a model can look fair on sex alone and fair on age alone yet fail specifically for older women, because the harm lives at the crossing of two axes that each look fine in isolation. Audit the intersections you have the power to measure, and treat missing intersections as unmeasured risk, not as evidence of fairness. Finally, remember that collecting demographic labels is itself a privacy-sensitive act governed by the biometric-data constraints of Chapter 30: you need enough attribute information to detect gaps without over-collecting sensitive data, a genuine tension to design around rather than ignore.

What's Next

In Section 26.7, we bring the whole chapter to bear on real deployments across fitness, healthcare, safety, and accessibility, the domains where taxonomy choices, sensor placement, segmentation, personalization, and the fairness gaps of this section stop being abstractions and start determining whether a product helps or fails the person wearing it.