Investigating the Predictive Potential of Large Language Models in Student Dropout Prediction

In the landscape of educational analytics, the usage of Machine Learning (ML), Deep Learning (DL), and Survival Analysis (SA), for student dropout prediction often encounters challenges in effectively detecting dropout cases and explaining dropout reasons. This is due to many challenges such as data imbalance, data processing issues, cold start problem, and the limited explainability of the predictive models. This paper explores the usage of Large Language Models (LLMs) for dropout prediction to tackle the previous challenges. We introduce an approach that leverages the adaptability and contextual understanding of LLMs to discern subtle indicators of potential dropout risks. In particular, we employ a Retrieval Augmented Generation (RAG)-assisted few-shot learning paradigm and prompt engineering to transfer the knowledge of LLMs. An intensive experimentation of our approach has been conducted using real-life Moroccan student data containing academic, demographic, and socio-economic information, to predict yearly school dropouts. Our findings highlight that LLMs outperform baseline ML models while showing the ability to produce textual analysis of students’ data. Thus, LLMs have a promising potential to be employed as student dropout prediction assistants in educational institutions hoping to mitigate this phenomenon.