線上教育平台的蓬勃發展，使得傳統教育型態進行轉變，隨著許多教育機構開始採用線上教育平台提供學生教育資源，如何確保學生學習品質的議題也日益重要。為了幫助授課教師掌握學生學習狀況與及時提供介入輔導，已有許多研究學者引入教育數據挖掘(Educational data mining ,EDM)於教育環境，透過機器學習、統計學，對學生之學習行為進行探索並對學習成效進行預測。因此，本研究透過分析學習行為，建立有效準確的學習成效預測模型，期望能確保線上教育平台的教育品質。 本研究收集了三個不同微積分班級的學習歷程，包含Open edX平台、MapleTA平台的線上行為與線下的實體作業與測驗。相關研究指出，一般將預測學習成效歸類成兩種問題，分別是迴歸與分類，因此本研究分別對這兩種作法進行探討，並提出ensemble method與常見的演算法進行比較，證明ensemble method能更進一步提升預測準確度。在迴歸問題方面，本研究先從常見的六種迴歸演算法中找出一種較穩定、準確的演算法，並以此演算法為基礎加入資料點分類與指標函數(indicator variables)；另外，在分類問題方面，本研究透過在建立分類模型的過程中加入重採樣(resampling)與投票機制(voting)來解決原始資料集中資料不平衡與單一演算法預測效能不足的問題，最後將此兩種ensemble method實作於三個微積分班級上，證明了ensemble method確實有達到改善的效果。 ;The rapid development of the online education platform has changed the traditional education. With the adoption of Massive Open Online Course(MOOCs) in many education institutions, the issue of ensuring the quality of student learning becomes more and more important. In order to help instructors keep track of the progress of students and provide interventions to at-risk students, many researchers have introduced Educational Data Mining (EDM) into educational environment and apply machine learning and statistics not only to explore students’ learning behaviors but also to predict student academic performance. Therefore, this study analyzes learning behavior and build an effective and accurate predictive model of predicting student academic performance. Real data was collected from three MOOCs and MapleTA enabled calculus course, which comprise of video viewing behavior, online assessment behavior, homework score and exam score. Many researchers generally predicting student academic performance by applying regression algorithms and classification algorithms. Therefore, this study explores these two approaches separately and proposes ensemble methods that are better than common algorithms. In terms of regression, this study first finds a relatively stable and accurate algorithm from the common six regression algorithms, and improve this algorithm through applying classifier to assign indicator variables to data points. In terms of classification, we add resampling technology and voting classifier to solve unbalanced data problem and bad performance by using single-algorithm. Finally, the two ensemble methods are implemented on three calculus classes, demonstrating that the ensemble method does achieve an improvement.
