ACT_预测大学新生GPA：传统和公平意识机器学习模型的比较研究

Research Report October 2025 Predicting College Freshman GPA: A Comparative Study of Traditional and Fairness-Aware Machine Learning Models EDGAR I. SANCHEZ © 2025 by ACT Education Corp. All rights reserved. | R2503ACT Research | Research Report | October 2025 2 © 2025 by ACT Education Corp. All rights reserved. | R2503 Conclusions This study concludes that traditional logistic regression models, particularly those using ACT Composite scores, tend to demonstrate better fairness metrics across subgroups compared to a fairness-aware machine learning gradient-boosted machine model. The exclusion of race/ethnicity from predictive models does not introduce notable bias and may even enhance fairness, providing a lawful and effective way to evaluate students’ potential success in college. The findings suggest that postsecondary institutions should adopt a combined approach using both high school GPA and ACT scores to strike a balance between fairness and predictive accuracy, while being cautious with fairness-aware machine learning models due to their complexity and potential biases. So What? The practical importance of this study lies in its implications for postsecondary institutions, especially in light of the 2023 U.S. Supreme Court decision that ended affirmative action in college admissions. By comparing traditional logistic regression models with fairness-aware machine learning models, the study provides insights into how institutions can develop predictive models that balance fairness and accuracy without relying on race/ethnicity. This is crucial for complying with legal mandates while promoting equitable educational outcomes. The findings suggest that using a combined approach of high school GPA and ACT scores can help promote fairness and improve the predictive accuracy of student success, allowing institutions to more effectively allocate resources and supports to students who need them most. Now What? First, postsecondary institutions should consider adopting a combined approach using both high school GPA and ACT scores to develop predictive models that balance fairness and accuracy. This approach helps mitigate potential biases that arise when a model relies solely on one metric, particularly for African American and low-income students. Additionally, institutions should explore the use of fairness-aware machine learning models, but with caution, as these models may require further optimization to address potential biases and may be difficult to justify to parents and lawmakers. Postsecondary institutions should also focus on transparency and accountability in their decision-making processes, ensuring that the selection of specific models and criteria can be easily understood by and justified to students, parents, and legal authorities. Finally, institutions should incorporate nontraditional factors such as personal essays, socioeconomic background, and school context into their admissions processes to promote a more holistic evalua