基於學習行為特徵之學習者電腦中介溝通能力預測模型設計-以網路合作問題導向學習為例

本研究旨在透過發展之學習行為微歷程記錄器蒐集網路合作式問題導向學習課程活動中學習者的學習行為微歷程，根據所蒐集資料歸納學習行為特徵分類架構，發展基於學習行為微歷程特徵建立電腦中介溝通能力預測模型的有效方法，以建立即時預測模型並檢驗預測模型正確率與穩定性。此外，也探討學習情境與學習者特質對於電腦中介溝通能力預測準確度之影響。研究結果顯示，運用線性迴歸、樹狀迴歸演算法、序列最小優化迴歸演算法、卷積類神經網路演算法以及多層感知機演算法建立之預測模型都具有良好的預測能力，其中序列最小優化迴歸演算法建模具有最高的預測準確率與穩定性，其預測平均絕對誤差低至0.2522分。而「溝通行為」與「溝通目的」為影響電腦中介溝通能力預測模型之關鍵特徵子群，並且發展電腦中介溝通能力預測模型必須考量學習情境及活動與學習者學習行為微歷程的關係。若學習者有更多的互動討論行為、較熟悉教學活動主題、具有較多的學習者案例，都有助於建立預測成效更準確的電腦中介溝通能力預測模型。

關鍵字

電腦中介溝通；行為特徵；微歷程；機器學習；預測模型

並列摘要

This study aims to develop a Computer-Mediated Communication (CMC) competence forecasting model based on several considered well-known machine learning schemes and learning behavior features collected by a micro-behavior recorder from the learners while using an online collaborative problem-based learning system to perform a problem-solving learning activity. To summarize the big data generated from a huge amount of micro behaviors into the useful behavior features for constructing a good CMC competence forecasting model, this study developed the learning micro-behavior classification structure according to the collected data features and the concept of CMC. An effective method for constructing a high correctness and stableness CMC competence forecasting model was proposed and examined. In addition, the effects of learning situations and learners’ personal trait on the accuracy of CMC competence forecasting model were also discussed. The results show that the CMC competence forecasting models developed by the linear regression algorithm, M5P algorithm, sequence minimum optimization regression algorithm, convolutional neural network algorithm and multi-layer perceptron algorithm all have good prediction performance. Among the five machine learning schemes, this study found that the sequence minimum optimization regression algorithm has the highest prediction accuracy and stability of prediction accuracy. The key features that influence the forecasting accuracy most are “communication behavior” and “communication purpose.” Moreover, the development of the forecasting model must consider the relationship between learning situations and learners’ personal trait such as discussion trait and familiar degree with the problem-solving subject because those would affect the accuracy of the developed forecasting model.