FairML estimates how strongly a predictive model depends on its input features by changing inputs and observing the model’s predictions. That can help investigate a model’s behavior, including in a fairness assessment, but feature dependence alone does not determine whether a system is fair.
What FairML measures
FairML is a Python toolbox for auditing predictive models when their internal workings are unavailable or are treated as a black box. The project describes it as an end-to-end toolbox that quantifies the relative significance of model inputs using model compression and four input-ranking algorithms (FairML on PyPI).
In practical terms, FairML estimates relative feature dependence: it asks how predictions change when input attributes are varied. The resulting ranking indicates which inputs appear more influential to the audited model under the chosen data and procedure. It is not, by itself, a measure of whether a model is accurate, unbiased, lawful, or fair.
How the audit works
Change inputs and observe predictions
The method perturbs model inputs and records how the predictions respond. The project’s demo accepts a black-box function and sample data in a pandas DataFrame without missing values. The sample should represent cases the model is expected to encounter, because the audit’s evidence is tied to the inputs it examines. The demo returns a dictionary of feature-dependence results across repeated runs (FairML on PyPI).
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
#1 Best Overall
- Use scikit-learn to track an example ML project end to end
- Explore several models, including support vector machines, decision trees, random forests, and ensemble methods
- Exploit unsupervised learning techniques such as dimensionality reduction, clustering, and anomaly detection
- Dive into neural net architectures, including convolutional nets, recurrent nets, generative adversarial networks, autoencoders, diffusion models, and transformers
- Use TensorFlow and Keras to build and train neural nets for computer vision, natural language processing, generative models, and deep reinforcement learning
The 2017 explanation says the approach can be used with a classifier or regressor that provides a predict function (Fast Forward Labs’ FairML explainer). This describes the interface in that account; it does not establish that the package works with every current model framework or Python environment.
Account for correlated features
Correlated inputs complicate feature ranking: changing one feature may indirectly stand in for information carried by another. FairML uses orthogonal projection to remove linear dependence between attributes during perturbation. The explainer also describes basis expansion and a greedy search over expansions as a way to address nonlinear dependencies. Linear projection alone would not capture nonlinear relationships, and results still depend on how well the procedure represents the data’s dependencies (Fast Forward Labs’ FairML explainer).
Rank #2
What FairML can—and cannot—say about fairness
A feature-dependence ranking can surface attributes that merit scrutiny, but it does not settle a fairness question. Fairness depends on the setting and the definition being applied. A ranking does not show on its own whether a feature is an appropriate predictor, whether a protected group experiences a particular error rate, or whether a system meets a legal or policy standard.
Use the result as one piece of an assessment: interpret it alongside the model’s purpose, the population and data represented in the audit, relevant outcomes, and an explicitly chosen fairness criterion. Do not treat a high ranking as proof of discrimination or a low ranking as proof that a model is fair.
What the COMPAS example actually audited
The FairML article discusses COMPAS risk scores using data collected by ProPublica about defendants in Broward County, Florida. Because COMPAS was proprietary, the demonstration did not query the COMPAS algorithm directly. It trained a logistic-regression proxy from the collected attributes and treated that proxy as a reasonable approximation for the demonstration (Fast Forward Labs’ FairML explainer).
In that proxy-model audit, prior offenses ranked highest, followed by the African American attribute. The article reports that accounting for multicollinearity strengthened the apparent association with that attribute. These are results about the proxy model and its audit, not a direct FairML audit of COMPAS itself.
Rank #4
Separately, the article quotes ProPublica’s analysis of a sample of about 7,000 people in Broward County. ProPublica reported that COMPAS “correctly predicts recidivism 61 percent of the time” and that Black defendants were “almost twice as likely as whites to be labeled a higher risk but not actually re-offend” (ProPublica’s analysis). The 61 percent figure and the disparity claim are ProPublica’s findings as relayed in the article; they are not performance results from FairML or the logistic-regression proxy. The disparity statement specifically concerns false high-risk labels.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How FairML differs from related tools
Tools for model interpretability and fairness answer different questions. ACM FAccT’s directory lists FairML alongside LIME and Aequitas, describing LIME as a tool for explaining individual predictions and Aequitas as an open-source bias-audit toolkit (ACM FAccT tools directory).
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesBest Value
| Tool | Question it is presented as answering | What that framing does not establish |
|---|---|---|
| FairML | How does the model depend relatively on its inputs? | A feature ranking alone does not establish fairness or provide a direct audit of a proprietary model that was not queried. |
| LIME | How can an individual prediction be explained? | The directory does not provide a current feature-by-feature comparison or establish that it answers FairML’s global-ranking question. |
| Aequitas | How can a model be examined through a bias audit? | The directory does not provide a current performance benchmark against FairML. |
This is a distinction in audit purpose, not a claim that one tool is better. The directory is not a current benchmark, so it does not support a performance ranking among them.
Is FairML current software?
PyPI records the FairML release date as June 28, 2017 (FairML on PyPI). That date establishes when the listed release appeared, but the available project information does not establish present-day maintenance or compatibility with current Python dependencies. Before relying on it, check the package’s current installation requirements and test it in the environment and model workflow you intend to use; do not assume the historical demo is production-ready today.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
