Treffer: A Bias-Aware Deep Learning Framework for Hierarchical Microcredential Classification

The growing demand for microcredentials in education and workforce development necessitates scalable, accurate, and fair assessment systems for both soft and hard skills based on students' lived experience narratives. Existing approaches struggle with the complexities of hierarchical credentialing and the mitigation of algorithmic bias related to gender and ethnicity. In this paper, we propose a novel deep learning framework that integrates hierarchical classification based on dynamic thresholding with a dual deep Q network dueling (DDQN dueling) for bias mitigation. Our method improves predictive performance at all three levels of microcredential classification, achieving an increase in 7% sensitivity and an improvement in 6% specificity over baseline models. Furthermore, our framework significantly improves fairness by reducing gender and ethnicity bias, as measured by equalized odds, by over 20% compared to conventional approaches. Extensive evaluations on a dataset of 3,000 student narratives demonstrate a 12% improvement in the F1 score and a 5% increase in AUROC relative to existing methods. These results underscore the effectiveness of our approach in advancing both hierarchical classification accuracy and fairness in real-world educational applications. [For the complete proceedings, see ED675583.]

As Provided