本論文提出了一種創新的方法，結合自然語言處理（NLP）、嵌入技術及生成式人工智慧（Generative AI）干預，旨在預測並提升線上教育中的學習表現。研究的第一階段專注於構建一個強大的學習表現預測模型，該模型整合了來自 BookRoll 和 Viscode 系統的數據。通過分析學習行為日誌並將其轉化為自然語言描述，使用 BERT 等嵌入技術和聚類演算法識別學習行為中的模式。這些洞見被結構化為學習行為，並作為深度學習模型的輸入，達到高達95%的預測準確率顯著超越傳統方法，能夠捕捉學生行為的時間性與語義特徵，並提供卓越的解釋性與擴展性。我們識別出一組共通學習行為（Common Learning Behaviors, CLB），這些行為與高表現高度相關，並可作為設計個人化介入策略的基準。研究的第二階段著重於在真實教育情境中實施個性化干預。以反思性思維及自我調節學習（Self-Regulated Learning, SRL）理論為基礎，生成式人工智慧（Generative AI）提供學習行為的個性化洞察，支持學生進行自我反思。透過指導性的反思問題及動態反饋機制，學生能夠採取適應性的學習策略，提升時間管理與自我評估的能力。實驗結果顯示，AI 輔助的干預顯著改善了學生與 CLBs 的行為一致性，縮短行為距離，並促進有意義的自我調節實踐。特別是，時間管理被證實是連結自我評估與學術表現的關鍵橋樑。本研究不僅推動了學習表現預測方法的進步，還展現了人工智慧驅動的個性化干預在促進深度學習體驗中的潛力。透過結合尖端技術與教育理論，本研究提供了一個轉型性框架，用於提升自我調節學習及學術成果。研究結果強調了適應性干預的重要性，為未來在個性化與可擴展教育系統中的探索鋪平了道路。;This dissertation introduces an innovative approach to predicting and enhancing learning performance in online education by leveraging Natural Language Processing (NLP), embedding techniques, and generative AI-based interventions. The first phase focuses on a robust performance prediction model that integrates data from BookRoll and Viscode systems. By analyzing behavioral logs and converting them into natural language descriptions, embeddings such as BERT and clustering algorithms identify patterns within learning behaviors. These insights are structured as learning behavior and serve as inputs for a deep learning model, achieving a prediction accuracy of up to 95%. We identify the Common Learning Behaviors (CLB) which can achieve high performance and can be the baseline for personalized intervention. This model outperforms traditional methods by capturing the temporal and semantic dimensions of student behaviors, offering unparalleled interpretability and scalability. The second phase of the study involves implementing personalized interventions in real educational scenarios. Grounded in reflective thinking and self-regulated learning (SRL) theories, generative AI (GenAI) supports self-reflection by providing tailored insights into learning behaviors. Through guided reflective prompts and dynamic feedback loops, students engage in adaptive learning strategies, enhancing their time management and self-evaluation capabilities. Experimental results demonstrate that AI-assisted interventions significantly improve alignment with CLBs, reduce behavior distances, and foster meaningful self-regulation practices. Notably, time management emerges as a critical bridge between self-evaluation and academic performance. This research not only advances the methodology of performance prediction but also highlights the potential of AI-driven personalization in fostering deeper learning experiences. By combining state-of-the-art technologies with educational theories, the study provides a transformative framework for enhancing self-regulated learning and academic outcomes. The findings underscore the significance of adaptive interventions, paving the way for future explorations in personalized and scalable education systems.
