Modelling the long-term fairness dynamics of data-driven targeted help on job seekers

Area of Science:

• Computational social science
• Labor economics
• Algorithmic fairness

Background:

• Public agencies increasingly use data-driven decision support systems.
• These systems can lead to ethical concerns and negatively impact marginalized groups.
• Existing models often lack long-term dynamic assessments of fairness.

Purpose of the Study:

• To assess the long-term fairness implications of data-driven labor market interventions.
• To compare intervention models with and without the use of protected attributes (e.g., group affiliation).
• To investigate the trade-offs between different fairness goals over time.

Main Methods:

• Combined statistical, data-driven, and dynamical modeling approaches.
• Developed a model simulating a public employment authority using a data-driven intervention model.
• Analyzed model dynamics, focusing on fairness metrics and the impact of using group affiliation as a predictive feature.

Main Results:

• Intervention models using group affiliation can increase predictive accuracy but may exacerbate fairness issues.
• Long-term fairness effects depend on the complex dynamics of the labor market.
• Quantifying trade-offs between fairness goals requires careful consideration of the entire system.

Conclusions:

• Dynamical modeling is essential for evaluating the long-term fairness of algorithmic decision-making in public resource allocation.
• Ignoring labor market complexities can lead to unintended negative consequences for disadvantaged groups.
• Future interventions must incorporate robust fairness assessments and dynamic modeling to ensure equitable outcomes.