| 摘要: | STEM程式已成為國內大專院校必修課程，而問題解決教學法是適合培養學員的計算思維能力。本研究使用聊天機器人輔助程式教學系統與問題解決教學法融入STEM程式課程來提升學習效益，探討學生的學習成效、學習感受以及學習行為之間差異。
實驗對象為北部大專院校學生，實驗主題為程式設計課程。實驗活動之後，蒐集測驗、問卷、訪談以及系統平台的操作紀錄等數據。依據前測將受試者分為三組進行比較學習成效差異，分析學習感受與學習行為間的相關性。接著，使用機器學習技術分析學習感受與學習行為對答題驗證的結果。
實驗結果發現大部分學生對於程式設計有一定程度的背景知識，部分學生注重實作能力，並不在意理論知識，所以成績提升有限；低先備與高先備知識組認知過程相似，學習感受問卷的成就、組織、參考等向度與學習行為有顯著的相關性，成就向度分數高的會重複觀看題目並提交程式碼驗證，組織向度分數高的會嘗試不同解題方法，參考向度分數高的課程隨堂測驗表現較好；在此研究中，監督式學習中SVM表現最好，其F1 score為0.758。使用與學習感受問卷向度相關性高的特徵進行分類後，其F1 score為0.74，這表示說可以藉由去除相關性較低的特徵實現資料降維，發現最好的STEM程式課程的教育資料探勘方法。
;STEM programming education has become a required course in domestic universities, and problem-solving instructional methods are suitable for cultivating students′ computational thinking ability. This study integrates chatbot-assisted instructional systems and problem-solving instructional methods into the STEM programming course to enhance learning performance, and we investigate the differences in students′ learning effectiveness, perceptions, and behaviors.
The experimental subjects are college students in Northern Taiwan, and the theme of the experiment is the STEM programming course. After the experiment, we collected data from tests, questionnaires, interviews, and system logs. According to the pre-test, the students were divided into three groups to compare the differences in learning effectiveness and analyzed the correlation between learning perceptions and behaviors. Next, we analyze the learning perceptions and learning behaviors using machine learning techniques to compare the results of verifying answers.
Results show that most students have a certain degree of background knowledge in STEM programming. Moreover, some students focus on practical abilities and do not care about theoretical knowledge, so their programming performance is limited. The cognitive process of low and high prior knowledge groups is similar. In the learning perception questionnaire, the effort, organization, and reference dimensions of students are significantly correlated with students′ learning behaviors. The high effort score of students can repeatedly improve the programming questions and submit the answer several times for validation. The high organization score of students can try different solutions, and the high reference score of students can get better programming performance in quizzes. In this study, the SVM performs the best in all machine learning algorithms, and the F1 score of SVM is 0.758. After classifying the highly correlated features on the learning perception questionnaire, the F1 score of SVM is 0.74. This finding indicated that the ideal dimension reduction is realized by removing low correlated features on learning perceptions. Finally, we find the best educational data mining approach in STEM programming courses. |
