Research Focus

Fairness & Bias

Algorithmic fairness, bias detection, and preservation of human decision-making

What is Fairness in AI?

AI systems can perpetuate and amplify biases from historical data, disproportionately affecting marginalized groups. Fairness (equity) seeks to ensure that automated decisions are just, impartial, and non-discriminatory.

Our work integrates bias detection, fairness metrics and mechanisms that preserve human autonomy and dignity against fully automated decisions.

Dimensions of Equity

Different perspectives on what "fair" means in AI

Demographic

Demographic Parity

The positive prediction rate should be equal among all demographic groups. Prevents a model from systematically favoring certain groups.

Individual

Individual Fairness

"Similar" individuals receive similar treatment. Based on similarity of relevant features, not group membership.

Balance

Calibration

Predicted probabilities are accurate for all groups. If the model predicts 70% for a group, approximately 70% are true positives.

Mitigation Strategies

Techniques to reduce bias in AI systems

Pre-processing

Balance training data, resample by demographic group, or anonymize sensitive features before training.

In-processing

Include fairness constraints in the objective function during training. Trade-off between accuracy and fairness controlled.

Post-processing

Adjust predictions after training to ensure equity. Useful when you cannot modify the original model.

Current Challenges

Open problems in fairness and human autonomy

Conflict Between Definitions

There is no universal definition of "fair". Different fairness metrics can be mutually mathematically incompatible.

Detection of Implicit Bias

Bias can be indirectly encoded in non-demographic features. We develop methods to identify them.

Human Autonomy vs Efficiency

Preserving human control in critical decisions can reduce efficiency. We seek balance where humans oversee automated decisions.

Practical Applications

Domains where fairness and bias mitigation are essential

Hiring and Recruitment: Ensuring automated screening tools do not discriminate against gender, ethnicity, or age, and auditing algorithms for equitable candidate ranking.
Credit and Lending: Preventing historical biases from affecting loan approval decisions and verifying that scoring models treat all demographic groups fairly.
Healthcare: Guaranteeing that clinical decision-support systems provide equitable recommendations across diverse patient populations and socioeconomic backgrounds.
Education: Building adaptive learning platforms that avoid reinforcing stereotypes and ensure equal access to resources for all students.
Law Enforcement: Auditing predictive policing and risk-assessment tools to prevent discriminatory profiling and protect civil rights.
Public Administration: Ensuring fair distribution of social benefits and services through transparent, bias-free automated allocation systems.

Our Research

In the Fairness and Bias line of the Trustworthy AI group at IAFER, we develop methods to detect, measure, and mitigate bias in AI systems, ensuring that automated decisions treat all groups equitably.

🚀 Active Research Lines

Demographic bias detection: Methods to identify disparities in predictions across population groups
Pre, in and post-processing mitigation: Techniques at different stages of the ML pipeline to correct inequities
Fairness-utility trade-off: Methods that balance overall performance with fairness across groups
Preservation of human autonomy: Ensuring human supervision in critical decisions that affect individual rights

📚 Publicaciones

Loading publications...