Scholarlink

<p>Personalized learning in visual design education demands AI tutors capable of deep, adaptive scaffolding grounded in domain-specific pedagogy—far beyond the capacity of generic conversational systems. Existing intelligent tutoring systems commonly lack a principled computational model that coherently integrates structured domain knowledge with long-term learner progression. To close this gap, we introduce the <b>Adaptive AI Tutor for Visual Design (AAT-VD)</b>, a unified architecture that tightly couples a hybrid neural recommendation engine, a Markov Decision Process (MDP)-based learning path planner, and an explainability framework built on attention visualization and SHAP attribution. The system encodes visual design domain knowledge in a prerequisite-aware Knowledge Graph (VG), dynamically tracks each learner through a composite state vector encoding knowledge proficiency, stylistic preferences, and recent interaction context, and scores candidate content items with a Bayesian Personalized Ranking (BPR)-trained scoring function. The MDP planner optimizes cumulative discounted learning gain via REINFORCE with a learned baseline, enabling coherent multi-step pedagogical sequencing rather than greedy, myopic selection. In a pre-registered, between-subjects controlled experiment with 80 participants, AAT-VD outperforms three baselines—a conversational tutor, a fixed curriculum, and a popularity-based recommender—achieving a <b>21.7% learning gain</b> (versus 15.2% for the best-performing baseline, the Original AI Tutor), a learning efficiency of <b>0.48 gain/min</b>, Precision@5 of <b>0.68</b>, and mean Perceived Usefulness of <b>4.32/5.0</b>. A comprehensive ablation study isolates the contribution of each component, confirming that the path planner and knowledge graph provide statistically significant independent gains in learning outcomes, while stylistic preference modeling significantly enhances creative engagement. These results establish AAT-VD as a principled, end-to-end framework for adaptive creative education and provide replicable design guidance for the broader AI-in-education community.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AAT-VD: an Adaptive AI Tutor with a Computational Model for Visual Design Learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Xudong Ma

摘要

Personalized learning in visual design education demands AI tutors capable of deep, adaptive scaffolding grounded in domain-specific pedagogy—far beyond the capacity of generic conversational systems. Existing intelligent tutoring systems commonly lack a principled computational model that coherently integrates structured domain knowledge with long-term learner progression. To close this gap, we introduce the Adaptive AI Tutor for Visual Design (AAT-VD), a unified architecture that tightly couples a hybrid neural recommendation engine, a Markov Decision Process (MDP)-based learning path planner, and an explainability framework built on attention visualization and SHAP attribution. The system encodes visual design domain knowledge in a prerequisite-aware Knowledge Graph (VG), dynamically tracks each learner through a composite state vector encoding knowledge proficiency, stylistic preferences, and recent interaction context, and scores candidate content items with a Bayesian Personalized Ranking (BPR)-trained scoring function. The MDP planner optimizes cumulative discounted learning gain via REINFORCE with a learned baseline, enabling coherent multi-step pedagogical sequencing rather than greedy, myopic selection. In a pre-registered, between-subjects controlled experiment with 80 participants, AAT-VD outperforms three baselines—a conversational tutor, a fixed curriculum, and a popularity-based recommender—achieving a 21.7% learning gain (versus 15.2% for the best-performing baseline, the Original AI Tutor), a learning efficiency of 0.48 gain/min, Precision@5 of 0.68, and mean Perceived Usefulness of 4.32/5.0. A comprehensive ablation study isolates the contribution of each component, confirming that the path planner and knowledge graph provide statistically significant independent gains in learning outcomes, while stylistic preference modeling significantly enhances creative engagement. These results establish AAT-VD as a principled, end-to-end framework for adaptive creative education and provide replicable design guidance for the broader AI-in-education community.