近年程式設計能力逐漸成為教育界發展重點，然而程式語言結合運算思維及編程操作，高入門門檻往往導致學生之間學習進度差異過大使教學困難，評估學生學習狀況成為程式語言教育的重要課題。在此結合線上編程系統、系統日誌過濾整合技術萃取學習參與度，提供學習行為參與度即時視覺化呈現，提供教學端面板進行學習行為監控。 對於作業繳交原始碼，以程式編寫風格評估以及程式碼缺陷預測模型進行程式碼品質分析，提供評量依據以及自我檢查的資訊。為了節省教學負擔，透過過去NASA MDP專案經驗中的資料集作為訓練資料，取代教學端逐一檢視程式碼並且給予是否存在缺陷之人工標記，以遷移式學習中的場域對抗式類神經網路(Domain-Adversarial Neural Network)技術建立程式碼缺陷模型，而結果顯示DANN的場域適應性較傳統機器學習好，DANN模型之proxy A-distance (PAD)數值為0.9，傳統機器學習方法之PAD則都大於1.8。 最後，研究程式編寫風格與程式碼缺陷的關聯；透過皮爾森相關係數檢測兩組特徵，發現其低關聯(小於0.25的對)比例近七成，而透過多元線性回歸分析，以兩組指標不同特徵(features)輸入模型訓練，建立程式編寫風格分數預測模型，程式編寫風格指標模型pMSE值為9.349，而程式碼缺陷指標pMSE為13.686，透過只有0.7的平均預測誤差發現雖然程式碼編寫風格與程式碼缺陷特徵間沒有直接關聯卻對於程式編寫風格分數有相近的預測能力。;In recent years, programming skill have gradually become the focus of education development in the education sector. However, programming combines computational thinking and coding operations. High learning thresholds often lead to difficult teaching due to the huge differences in student learning progress. Measure students’ learning becomes an important issue in programming education. We combination of online programming system, system log filter integration technology to extract learning participation, provide immediate visualization of learning behavior engagement to help teachers to monitor students’ learning behavior. For assignment of the source code, code quality analyzes by coding style measure and code defects prediction model to provide evaluation and self-examination. In order to reduce the burden of teachers, we use NASA MDP project experience as a training data, instead of the teaching side to view lots pieces of code and give the artificial mark of whether the code is defect or not. And use Domain-Adversarial Neural Network to build code defects prediction model. The result shows that DANN′s domain adaptability is better than traditional machine learning. The DANN model’s proxy A-distance (PAD) value in 0.9, and traditional machine learning methods have a PAD greater than 1.8. Finally, we study the relationship between coding style and code defects. We use Pearson correlation coefficient to find relationship of two sets. We found that its low correlation pairs (coefficient smaller than 0.25) is nearly 70%. After, we use different features and establish multiple linear regression to predict score of coding style model. The model build by coding style features’ pMSE value is 9.349, and the code defects features’ pMSE is 13.686. Through the average prediction error of only 0.7 difference, it is found that although there is no direct correlation between the coding style’s and the code defects’ feature, but two datasets have similar ability to predict score of coding style.
