2026 ASEE Annual Conference & Exposition

Historically Black Colleges and Universities (HBCUs) play a pivotal role in broadening access to higher education for underrepresented students in STEM disciplines. However, challenges such as lower retention and graduation rates underscore the need for advanced, data-driven educational interventions. Deep Knowledge Tracing (DKT) has emerged as a powerful approach for modeling students’ knowledge states and predicting learning outcomes, thereby enabling timely and targeted support. The study of this project advances the state of DKT research by addressing critical limitations in interpretability, low-resource environments, and real-world applicability, while also enhancing learning path recommendations to personalize STEM education at HBCUs.

HBCUs often face constraints in collecting and annotating large-scale educational data, which can hinder the performance of data-hungry deep models. To mitigate this, research on semi-supervised learning and ensemble DKT models demonstrates how unlabeled student interactions can be effectively leveraged to enhance model robustness in low-resource settings. In addition, the use of tabular generative AI to create synthetic student data has been shown to augment real datasets, improving learning outcome predictions. These innovations underscore the potential of generative and semi-supervised strategies to empower institutions with limited resources, ensuring that predictive models remain both effective and equitable. Moreover, a major barrier to deploying DKT in practice is its “black-box” nature, which limits trust among educators and students. To address this, we integrate SHAP (Shapley Additive Explanations) with DKT to quantify the influence of individual features, while leveraging large language models (LLMs) to generate natural-language explanations of model predictions. Another complementary effort combines ensemble learning with Item Response Theory (IRT) to improve interpretability in tracing student learning outcomes. These approaches not only improve predictive accuracy but also provide transparent, actionable insights into student abilities, item difficulty, and performance trends—bridging the gap between advanced analytics and classroom usability.

Beyond tracing student outcomes through advance DKT, this study also advances learning path recommendation by framing it as a multi-task sequence prediction problem. By sharing representations between DKT and recommendation tasks, and by introducing a non-repeat penalty to avoid redundant recommendations, the proposed multi-task model produces more effective, individualized learning trajectories.

Together, these contributions, improving interpretability, addressing data limitations, and refining recommendation strategies, lay the foundation for a more transparent, adaptive, and equitable learning ecosystem at HBCUs. Future research will continue to integrate cutting-edge AI techniques with educational theory to guide students more effectively and ultimately strengthen STEM education pathways for underrepresented populations. This work is supported by the NSF Improving Undergraduate STEM Education (IUSE) program.