An explainable learning analytics approach for AI-supported early warning in higher education using frequent risk automata
Year: 2026
Type: article
Abstract: AI-supported early warning in higher education is increasingly expected to help teachers and learning support staff identify students who may need timely attention. However, many prediction models provide probability scores or labels without sufficiently explicit behavioral evidence for learning-support interpretation, and high-recall settings may generate unnecessary false alarms. This paper proposes FRA-Learn as an explainable learning analytics approach for early warning. FRA-Learn transforms process-based learning records into temporal risk transactions and safety counter-evidence transactions, uses contrastive frequent-pattern induction to identify discriminative behavior combinations, and organizes risk and safety counter-evidence through a risk–safety competitive automaton that balances warning-supporting evidence with safety counter-evidence. The output includes warning labels, warning scores, activated risk patterns, and activated safety counter-evidence patterns as model-side interpretable evidence. Experiments on OULAD and KDD2015 show that FRA-Learn achieves strong Precision, F1 score, and AUC with a low false alarm rate (FAR) on OULAD, and maintains strong F1 while reducing FAR relative to high-recall baselines on KDD2015. The evaluation focuses on model-side interpretable evidence in public benchmarks; teacher-side use, classroom deployment and intervention effects require future validation.
Source: Discover Computing
Authors: Xin Dai, Chenjiao Liu, Zhi Yang, Xin Zhong, Yuyu He
Language: English
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: gold
Estimated APC paid: $1,520
OpenAlex ID: W7235176361
