Real-time detection of low-achieving groups in face-to-face computer-supported collaborative learning

This study investigates the feasibility of detecting low-achieving groups during face-to-face computer-supported collaborative learning in real-time. We collected in-class online activity data that records students' learning behaviors during face-to-face classes, and built prediction models that identify the at-risk groups at every minute during a class. A total of 88 pre-service teachers (56 female, 32 male) were recruited and assigned to 22 collaborative learning groups. The groups participated in two face-to-face collaborative argumentation classes that took place once a week over two consecutive weeks. The participants used online collaboration software, Trello, that allowed in-class online activity data collection during the classes. Ten group activity features were extracted from the data in three categories: participation, interaction, and quality of argumentation. Random forest algorithm was used to build the prediction models based on the group activity features. The results show that the models can detect the low-achieving groups with high accuracy even just a few minutes after the class begins. As the class progressed, the accuracy was improved. Additionally, the model identified the important group activity features that contributed to the group achievement in each phase of class. The results indicate that prediction models using in-class activity data can help instructors accurately identify at-risk groups in real-time and provide appropriate instructional support. An early warning system should be beneficial as well.