TL;DR
This paper introduces a framework to improve the accuracy and fairness of human-led decision-making—such as judicial bail or police stops—by optimizing how cases are assigned to experts. By modeling experts as threshold-based decision-makers and using sequential matching algorithms, the researchers demonstrate that we can mitigate individual biases and improve collective outcomes, even when the experts' specific biases are initially unknown.
What Problem Does This Paper Address?
In many critical fields like law, medicine, and academia, decisions are made by human experts. A judge decides on bail; a police officer decides on a stop-and-frisk; a professor decides on a paper's acceptance. Typically, these cases are assigned to experts randomly.
However, humans are fallible. We suffer from implicit biases, limited experience, and inconsistent reasoning. While a lot of research focuses on making algorithms fair, this paper asks: Can we make human systems fairer by changing how we assign tasks? If we know one judge is "harsh" and another is "lenient," can we route cases to them in a way that maximizes the accuracy of the whole system while ensuring groups (e.g., based on race) are treated equitably?
Methods
The authors model human experts as threshold decision-makers. An expert observes an individual's features and a "sensitive attribute" (like race) and makes a binary decision based on their internal threshold.
The core of their approach involves:
- Weighted Bipartite Matching: The problem of assigning $m$ individuals to $n$ experts is cast as a matching problem. Each "edge" in the graph represents the expected utility (accuracy) of an expert making a specific decision.
- Fairness Constraints: To ensure fairness (e.g., avoiding disparate impact), they add constraints to the matching. This transforms the task into a "bounded color matching" problem, which they solve using a bi-criteria algorithm that provides a 1/2 approximation guarantee.
- Posterior Sampling (Thompson Sampling): In the real world, we don't know an expert's internal bias threshold. The authors use a Bayesian approach: they maintain a probability distribution (a Beta prior) for each expert's threshold and update it as decisions are made. This allows the system to "explore" (learn about the expert) and "exploit" (assign cases to the best expert) simultaneously.
- Regret Analysis: They prove that using posterior sampling leads to "sublinear regret" ($O(\sqrt{T})$), meaning the system's performance converges to the optimal assignment over time. In contrast, simply "guessing" the threshold (point estimation) leads to "linear regret," where the system never stops making avoidable mistakes.
Datasets
The paper evaluates the algorithms using two primary sources:
- Synthetic Data: Generated using Bernoulli and Beta distributions to simulate a controlled environment where ground-truth biases and labels are known.
- COMPAS Recidivism Dataset: A well-known real-world dataset from ProPublica containing demographic features, criminal history, and risk scores for defendants in Broward County, Florida. Since the original data doesn't identify specific judges, the authors created "fictitious" judges with varying levels of bias to test the assignment logic.
Key Insights and Ideas
- Diversity is a Strength: Counter-intuitively, a diverse pool of experts (some harsh, some lenient) makes it easier to achieve fairness. If every expert has the exact same bias, the system has no "knobs" to turn. With a diverse pool, the algorithm can route cases to balance out individual biases.
- Matching as a "Fairness Layer": You don't necessarily need to retrain or "fix" the humans to get a fairer outcome. By treating the assignment process as a strategic layer, you can improve the system's collective behavior while respecting the experts' roles.
- The Necessity of Exploration: You cannot simply observe a few decisions and assume you know an expert's bias. The paper shows that active exploration (occasionally giving an expert a case that "tests" their threshold) is mathematically necessary to reach an optimal state.
Key Results
In both synthetic and real-world (COMPAS) tests, the proposed algorithms significantly outperformed random assignment in both accuracy (utility) and fairness (reducing disparate impact).
- Superiority over Randomness: The algorithm consistently achieves higher expected utility and lower disparate impact compared to random assignment baselines.
- Convergence: When thresholds were unknown, the algorithm's performance quickly approached the "Known Threshold" benchmark, confirming the effectiveness of the posterior sampling approach. This is reflected in the sublinear regret curves.
- Fairness Trade-offs: On the COMPAS data, the algorithm achieved fairness by strategically adjusting release rates—for example, by routing cases to ensure that the "benefit" (being released) was distributed more equitably across racial groups than it would be under random assignment.
Limitations and Future Work
- Uniform Prediction Ability: The current model assumes all experts are equally good at predicting the probability of an outcome (e.g., recidivism) and only differ in their thresholds (bias). In reality, some experts are simply better at their jobs than others.
- Static Experts: The paper assumes experts don't learn or change their biases over time. If a judge realizes they are being "routed" specific cases, they might shift their behavior.
- Ground Truth Requirement: The system requires a "true" conditional distribution (often provided by a machine learning model) to calculate the weights for the matching. If the underlying ML model is itself biased or inaccurate, the assignments will be too.
- Ethical Implementation: The authors acknowledge that assigning cases based on known human biases could be controversial or legally complex in sensitive fields like the judiciary.
Summary updated