Sequential Cohort Selection

Hortence Phalonne Nana, Christos Dimitrakakis

Abstract

We study the problem of fair cohort selection from an unknown population, with a focus on university admissions. We start with the one-shot setting, where the admission policy must be fixed in advance and remain transparent, before observing the actual applicant pool. In contrast, the sequential setting allows the policy to be updated across stages as new applicant data becomes available. This is achieved by optimizing admission policies using a population model, trained on data from previous admission cycles. We also study the fairness properties of the resulting policies in the one-shot setting, including meritocracy and group parity.

Introduction

In high-stakes selection processes such as university admissions or corporate hiring, decision-makers often face the challenge of setting policies before the actual applicant pool is known. This "one-shot" requirement demands policies that are both transparent and robust to shifts in population distributions. Traditional methods frequently assume a static, known population, failing to account for the "moving target" nature of real-world demographics or the fact that outcomes (e.g., academic performance) are only observed for accepted candidates.

This paper introduces a framework for Sequential Cohort Selection, designed to maximize expected utility while adhering to fairness constraints—specifically demographic parity and equality of opportunity. By updating policies across multiple cycles, the framework addresses the "cold start" problem and ensures long-term meritocracy.

Methodology

The selection process is formalized as an expected utility maximization task. The authors compare two primary policy structures:

  1. Linear Sigmoid Policy: A transparent model similar to logistic regression.
  2. Neural Network (NN) Policy: A flexible model capable of capturing non-linear relationships between applicant features (e.g., GPA, socio-demographics) and predicted success.

Utility is defined via a log-linear function that rewards the selection of high-performing individuals while applying a cost ($c$) per admission to prevent over-enrollment. To ensure fairness, the framework utilizes Expected Marginal Contribution (EMC) to measure an individual's value to the cohort. Fairness is enforced by penalizing discrepancies in selection rates between groups with similar EMCs (Equality of Opportunity) or different demographic backgrounds (Demographic Parity).

In the sequential setting, the system employs a Conditional Tabular GAN (CTGAN) to model the evolving applicant population, Bayesian Multivariate Linear Regression to predict outcomes under uncertainty, and Thompson Sampling to balance the exploration of new strategies with the exploitation of known success patterns.

Experimental Framework

Due to the interactive nature of selection policies, the study utilizes a simulator trained on a real-world dataset of Norwegian university admissions. This simulator generates synthetic applicant pools with features such as high school GPA and socio-demographic attributes, predicting academic outcomes across three courses to provide a realistic testing environment.

Results and Analysis

The Batch Size and Fairness Trade-off

A critical insight from the study is the "Batch Size Effect." Fairness violations are significantly more prevalent when the applicant pool is small. As the batch size increases, the law of large numbers allows the policy to stabilize. Neural network policies demonstrate a superior ability to reach near-zero fairness penalties much faster than linear models, particularly as the pool size grows.

Utility and fairness penalties across varying batch sizes for c=0.001 and c=0.1.

Adaptability in Sequential Settings

In multi-stage trials spanning ten years, the neural network policy exhibited remarkable stability and adaptability. While the logistic policy often remained rigid—struggling to achieve positive utility when admission costs were high ($c=0.1$)—the NN policy learned to become more selective over time. By prioritizing the highest-utility candidates, the NN policy maintained equitable representation while stabilizing at a utility above 12.

Evolution of utility and admission rates over a 10-year sequential selection period.

Conclusion

The findings suggest that flexibility is paramount for achieving fairness in complex selection environments. Neural network policies significantly outperform linear models when admission costs are high, as they can identify "pockets" of high-potential candidates that simpler models overlook. Furthermore, by pre-committing to a policy $\pi$ before observing the population, the framework preserves the transparency required by public institutions.

Future work will explore online learning for incremental policy updates and the "matching problem," where the model accounts for students choosing between multiple competing offers.

Summary updated

Abstract

We study the problem of fair cohort selection from an unknown population, with a focus on university admissions. We start with the one-shot setting, where the admission policy must be fixed in advance and remain transparent, before observing the actual applicant pool. In contrast, the sequential setting allows the policy to be updated across stages as new applicant data becomes available. This is achieved by optimizing admission policies using a population model, trained on data from previous admission cycles. We also study the fairness properties of the resulting policies in the one-shot setting, including meritocracy and group parity.
BibTeX
@misc{nana2025sequential,
  author = {Hortence Phalonne Nana and Christos Dimitrakakis},
  title = {{Sequential Cohort Selection}},
  year = {2025},
  eprint = {2508.16386},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.16386},
}
Figure