運用因果約束增強合理反事實解釋：以線上學習為例

Abstract

雖然線上教育中的預測模型能夠準確識別高風險學生，但若缺乏可行且合理的指導建議，這樣的識別仍顯不足。本論文提出 Causal-FACE，一個基於圖結構的反事實解釋框架，該框架整合了以有向無環圖（DAG）編碼的因果約束。我們使用開放大學學習分析資料集訓練 XGBoost 預測器，並透過結合 DirectLiNGAM 演算法、大型語言模型假設生成與人工驗證的混合程序來發現因果結構。解釋的產生是透過搜尋經驗 k-NN 圖，使用綜合邊成本來平衡鄰近性、局部資料流形密度以及因果合理性分數，並將產生的計畫轉化為學生可理解的明確行動建議。我們在相同可變性約束條件下，針對 535 名高風險學習者的匹配群組，將 Causal-FACE 與 FACE 及 DiCE 進行比較，並進行量化與質性評估，包括由人工專家小組支持的 LLM-as-a-Judge 評估。跨研究結果顯示，Causal-FACE 保持了流形真實性，提供更具因果一致性且更受信任的建議，同時保持相當的簡潔性，為虛擬學習環境中的可部署期中支援服務提供了方向。

While predictive models in online education can accurately identify at-risk students, this is insufficient without actionable, plausible guidance. This thesis introduces Causal FACE, a graph based recourse framework that integrates causal constraints encoded as a Directed Acyclic Graph (DAG). Using the Open University Learning Analytics Dataset, we train an XGBoost predictor and discover causal structure through a hybrid procedure that combines DirectLiNGAM, hypotheses from a large language model, and human validation. Explanations are generated by searching an empirical k-NN graph with a composite edge cost that balances proximity, local data manifold density, and a Causal Plausibility Score, and the resulting plans are verbalized into clear actions for students. We compare Causal FACE with FACE and DiCE on a matched cohort of 535 at risk learners under identical mutability constraints, and we conduct both quantitative and qualitative assessments, including an LLM-as-a-Judge review supported by a human panel. Across studies, Causal FACE preserves manifold realism and delivers recommendations that are more causally coherent and more trusted while remaining comparably concise, pointing to deployable mid course support in virtual learning environments.