Abstract
Algorithmic hiring systems must balance predictive accuracy with strict legal and ethical fairness standards. Traditional manual weighting by human experts is often suboptimal and prone to bias. This paper introduces a framework using Evolutionary Many-Objective Optimization (EMOO) and specialized Dirichlet-based genetic operators to automate the discovery of hiring algorithms. By treating fairness across multiple demographic groups as distinct objectives, the proposed method consistently outperforms human experts, identifying "glass-box" linear models that are both more accurate and fairer while remaining legally compliant.
Introduction
Hiring is a high-stakes balancing act where organizations seek to maximize candidate performance (predictive accuracy) while ensuring non-discrimination against protected groups (fairness). The industry standard remains "rational weighting," where human experts manually adjust the weights of various assessment scores based on heuristics. This manual process is slow, rarely identifies the optimal Pareto front—the set of solutions where fairness cannot be improved without sacrificing utility—and often struggles with complex legal constraints like the "four-fifths rule."
This research aims to automate this process using EMOO, ensuring that the resulting models are both interpretable and legally compliant by avoiding the use of protected attributes directly (disparate treatment) and minimizing adverse impact.
Methodology: Many-Objective Optimization and Dirichlet Operators
The authors frame hiring as a Many-Objective Optimization problem because fairness is not a monolithic goal; it requires minimizing adverse impact across multiple intersecting demographic groups (e.g., gender, race, and ethnicity) simultaneously.
Dirichlet-based Genetic Operators
A significant challenge in optimizing hiring weights is the constraint that weights must sum to 1.0. Standard evolutionary algorithms often use "repair" functions that can bias the search toward the center of the distribution. To address this, the authors proposed using Dirichlet distributions for sampling, crossover, and mutation. This ensures that every candidate solution naturally satisfies the sum-to-one constraint while maintaining a uniform search across the weight space.
Interpretability Constraints
To ensure results are explainable to stakeholders, the models were restricted to weighted linear combinations with specific rounding increments (e.g., weights as multiples of 0.05). This "glass-box" approach ensures that the logic behind a hiring decision is transparent and defensible.
Experimental Setup
The framework was evaluated on three real-world datasets from Fortune 500 companies:
- Leadership Dataset (N=377): 3 performance criteria, 3 diversity metrics.
- Sales Dataset (N=922): 3 performance criteria, 3 diversity metrics.
- Banker Dataset (N=416): 5 performance criteria, 3 diversity metrics.
Several state-of-the-art algorithms were tested, including NSGA-II, NSGA-III, and SPEA2-SDE, alongside a newly proposed model called TriGE.
Key Results and Insights
The "Expert" Ceiling
Experimental results demonstrated that human experts (Industrial/Organizational Psychologists) are significantly outperformed by EMOO. Humans struggle to mentally process the complex trade-offs between eight or more simultaneous objectives, whereas evolutionary algorithms can efficiently map the entire Pareto front.
The Dirichlet Advantage
Traditional crossover methods, such as Simulated Binary Crossover (SBX), tend to cluster solutions toward the middle of the search space. By using Dirichlet-based operators, the search remains diverse, which is crucial for finding the edges of the Pareto front where the most significant fairness/utility trade-offs occur.
Performance and Fairness Gains
In all datasets, NSGA-II and SPEA2-SDE consistently provided the best fit, achieving higher Hypervolume (HV) scores than manual rational weighting. The EMOO models identified algorithms that maintained the same level of predictive accuracy as human experts while significantly reducing adverse impact for female, Black, and Latinx applicants.
Conclusion
This research demonstrates that "black box" deep learning is not a prerequisite for fair hiring. Interpretable linear models, when optimized via EMOO with Dirichlet-based operators, provide superior performance and fairness compared to traditional human-led methods. Future work will focus on addressing larger sample sizes and intersectional fairness to ensure equitable outcomes for candidates belonging to multiple protected groups.
Summary updated