How Can Human Interaction Make University Admissions Fairer?

Artificial intelligences are increasingly used to decide university admissions. Yet, if the training data contain prejudices, the algorithms reproduce them, creating unfairness for vulnerable groups. To tackle this, researchers created an interactive method that lets users "tune" biases in real time.

The study started with six years of admission data from a large urban university. Several machine‑learning models were trained to predict acceptance, then examined for how they treated three sensitive groups: gender, ethnicity and first‑generation students.

The novelty lies in involving the user: through an interface, the operator can view bias metrics (e.g., acceptance rate disparity) and adjust training parameters. Each tweak produces a new, more balanced model according to the chosen fairness criteria. In three test scenarios, the system significantly reduced disparities without heavily sacrificing prediction accuracy.

The approach shows that transparency and human participation can make AI more trustworthy, especially in delicate fields like higher education. Ongoing interaction between experts and algorithms paves the way for more informed, less discriminatory decisions.

The proposed method blends supervised learning with multi‑objective optimization algorithms. Initially, the original model is trained on a labeled dataset (e.g., accepted or rejected). Then, disparity metrics such as disparate impact (DI) and equalized odds (EO) are computed for each sensitive variable. The user interface allows setting tolerance thresholds for these metrics. When a threshold is adjusted, the system applies a re‑weighting procedure: samples belonging to under‑represented groups receive higher weights, while those from over‑represented groups receive lower ones. This rebalancing is performed via constrained gradient descent, ensuring that the overall loss (accuracy) does not exceed a predefined limit. Moreover, the framework supports "counterfactual fairness": for each individual, a counterfactual profile (e.g., changing gender) is generated and the prediction is checked for invariance. The outcome is an iteratively refined model that converges to a Pareto front between accuracy and equity. Experiments showed DI reductions from 30% to 8% and EO improvements of 15%, with accuracy loss below 3%. The approach scales to larger datasets and can be integrated into existing admission pipelines, promoting more responsible AI governance in education.