Predicting Student Dropout Risk with Supervised Learning and Large Language Models: A Survey, a Hybrid Framework, and an Early-Warning Web Prototype

Abstract

The dropping out of students is a waste of the resources of learners and institutions. In Vietnam, early warning indications are documented in university grade vaults. Most early-warning research relies on supervised learning over structured records and so does not consider the behavioural and emotional signs present in unstructured student text such as advisor notes and learning feedback. Although large language models (LLMs) are capable of reading such text with semantic depth, this skill has not been rigorously linked to dropout prediction, let alone for Vietnamese. This research is a survey, framework and prototype study, not an empirically tested prediction model. It reviews dropout-prediction research in supervised, text-augmented, and LLM-based techniques, codes 32 validated sources along four dimensions, and anchors the synthesis in an anonymised dataset of around 1.22 million course-grade records from 35,763 students of Saigon University (2020–2025). Three notable findings are: tree ensemble methods continue to be the most deployable choice for structured records; class imbalance and leakage-safe temporal validation, not headline accuracy, determine practical value; and, LLM approaches are reported to improve minority-class detection where text exists, but remain untested for Vietnamese. Accordingly, the paper proposes a hybrid framework that fuses a Random Forest/XGBoost branch and a PhoBERT-based branch under explicit missing-modality handling, defines the dropout label, the student–semester prediction unit, and a leakage-safe evaluation protocol, and provides DropSense, a web prototype with transparent, illustrative weights. Future work includes a proposal for empirical validation.