Abstract:

The growing availability of online, distance education has expanded access to high-quality learning opportunities worldwide, playing an important role in advancing the United Nations’ Sustainable Development Goals related to inclusive and equitable education and reducing inequalities. However, student dropout remains a persistent challenge in this modality. For example, in Brazil, 65% of students in distance higher education dropped out. As a result, aiming to inform stakeholders like students, teachers, and administrative representatives, extensive research has focused on predicting student dropout using machine learning models. However, most of the existing approaches rely on black-box predictions or limited explainable artificial intelligence (XAI) outputs, providing little actionable value for key educational stakeholders such as instructors and institutional decision makers.

This paper evaluates an automated approach designed to transform model-level XAI insights into stakeholder-ready feedback aimed at mitigating dropout risks in online distance education. The proposed pipeline integrates predictive analytics with large language models (LLMs) to convert technical explanations into structured and actionable reports. First, a dropout prediction model was trained on a large-scale educational dataset (N = 1,591,482). Model explanations were then extracted using XAI techniques and provided as input to GPT-4o, which generates literature-grounded feedback reports tailored to three stakeholder profiles: students, professors, and institutional administration. The reports aim to translate predictive insights into concrete recommendations that support early intervention and targeted student support.

To assess the usability of the generated reports, we conducted an evaluation grounded in the Technology Acceptance Model. A total of 30 reports were purposefully selected and assessed by 20 university professors and 10 higher-education administrative representatives, resulting in a dataset of 300 individual evaluations. Mainly, quantitative results indicate high levels of intention to use across both stakeholder groups, suggesting strong acceptance of the approach as a decision-support tool. Exploratory analyses further reveal that perceptions of the reports vary according to evaluators’ professional background, such as their level of experience in higher education teaching and the subjects they teach. Moreover, complementary qualitative feedback highlights insights on how to improve the clarity, structure, and contextualization of the generated reports to better align with stakeholders’ operational needs.

Overall, the findings suggest that translating XAI insights into stakeholder-oriented narratives using LLMs represents a promising direction for addressing dropout in online distance education. By bridging the gap between predictive analytics and practical decision-making, the proposed approach can support instructors and institutional administrators in identifying at-risk students earlier, implementing targeted pedagogical interventions, and strengthening institutional strategies aimed at improving student persistence and minimizing educational disruption.

Keywords:

Higher Education, Dropout, Generative Artificial Intelligence, Technology Acceptance.