An Explainable and Personalized Generative AI Framework for Adaptive Learning, Intelligent Tutoring, and Student Performance Prediction

Abstract

Generative artificial intelligence has made it technically feasible to deliver individualised instruction at scale, yet three problems continue to limit its adoption in formal education. First, most deployed systems are generic: they respond to the prompt in front of them without any durable model of who the learner is, what the learner already knows, or where the learner is likely to fail next. Second, they are opaque: neither the learner nor the teacher can see why a particular hint, difficulty level, or risk score was produced, which corrodes trust and makes pedagogical accountability impossible. Third, predictive analytics and tutoring are usually built as separate systems, so a model that flags a student as at risk rarely does anything about it. This paper proposes EPGAI (Explainable and Personalised Generative AI), an integrated framework that unifies learner modelling, generative tutoring, and student performance prediction inside a single explainability-first architecture. EPGAI comprises five interacting layers: a multi-modal learner model that maintains cognitive, behavioural, and affective state; an attention-based knowledge tracing and performance prediction module that estimates mastery and forecasts assessment outcomes; a retrieval-augmented generative tutoring engine constrained by curriculum documents and pedagogical prompt templates; an explanation layer that converts SHAP attributions, counterfactual analysis, and rationale traces into classroom-appropriate natural language for three distinct audiences; and an adaptation policy that closes the loop by translating predictions and explanations into concrete instructional actions. The framework is evaluated on three public educational datasets (OULAD, ASSISTments, and EdNet) together with a mixed-methods study involving undergraduate learners and teachers. Results indicate that the integrated model improves predictive accuracy over conventional baselines, that retrieval grounding substantially reduces factually incorrect tutoring content, and that learner-facing explanations improve reported trust and self-regulation without harming learning outcomes. The paper also examines fairness across demographic subgroups, data-protection obligations, and the pedagogical risks of over-reliance on generative tutors, and concludes with a research agenda for classroom-scale validation.