FairLens - see the bias your model is hiding

• 

  thumbnail

• 

  fig_before_after_sex

• 

  fig_base_rates

• 

  fig_before_after_race

Inspiration

Machine-learning models increasingly decide who gets a loan, an interview, or a benefit — and they're judged on a single number: accuracy. But a model can be highly accurate and still systematically deny opportunities to women or to a racial group. The harm is invisible in the headline metric, so nobody notices. We wanted to make that invisible harm visible, measurable, and fixable — for anyone, not just ML researchers.

What it does

FairLens audits a real income classifier trained on the UCI Adult / Census Income dataset (48,842 people). It:

• trains a strong gradient-boosting model (87.6% accuracy, 0.93 ROC-AUC);
• audits it for bias by sex and race using legally-grounded fairness metrics (the U.S. EEOC four-fifths rule and equalized odds);
• shows it fails the four-fifths rule badly — a disparate-impact ratio of 0.32 by sex and 0.21 by race (anything under 0.80 is legal "adverse impact");
• applies a mitigation that closes most of the gap for ~1 accuracy point;
• wraps it in an interactive dashboard where you slide the decision threshold and toggle the fix to watch fairness change in real time.

How we built it

• Data cleaning (pandas): normalized three different spellings of "missing," recovered 3,620 hidden missing values, mode-imputed, and dropped a leak column (fnlwgt) and a redundant one (education).
• Modeling (scikit-learn): HistGradientBoostingClassifier on one-hot-encoded features, 70/30 stratified split.
• Fairness audit (Fairlearn): per-group selection / true-positive / false-positive rates, disparate-impact ratio, and equalized-odds difference.
• Mitigation (Fairlearn ThresholdOptimizer): group-aware thresholds for equalized odds — no retraining required.
• Dashboard (Streamlit + Plotly): live threshold slider, attribute selector, and a before/after comparison that shows the exact accuracy cost of fairness.

A fully-executed Jupyter notebook documents the entire data-science narrative.

Challenges we ran into

The sneakiest bug was in cleaning: calling .str.strip() before handling nulls silently turned every NaN into the literal string 'nan', hiding 3,620 missing values and making the data look clean when it wasn't — a good reminder that data-ethics work starts with honest data. Deployment surfaced another: the host provisioned Python 3.14, which had no prebuilt wheels for our pinned libraries, so we loosened them to wheel-available releases.

Accomplishments that we're proud of

Turning an abstract "AI ethics" concern into concrete, defensible numbers tied to actual U.S. anti-discrimination law (the four-fifths rule) — and shipping it as a tool a non-expert can actually operate.

What we learned

• Accuracy and fairness are independent; you have to measure fairness on purpose.
• "Fairness through unawareness" (just dropping sex/race) fails, because proxies like relationship and marital-status re-encode the protected attribute.
• There is no single definition of fairness — the honest move is to show the trade-off, not hide it.

What's next for FairLens

Upload-your-own-CSV auditing, more mitigation methods (reweighing, exponentiated gradient), and an exportable PDF "fairness report" for compliance teams.

Built With

• fairlearn
• jupyter
• matplotlib
• pandas
• plotly
• python
• scikit-learn
• streamlit