What Predicts Community in Virtual Computing Classrooms? A Machine Learning-Based Approach

The physical separation of people in online classes can produce feelings of isolation and disconnectedness in some students. Feelings of community - that is, a sense of belonging and that people in the community matter to each other - can help ameliorate these feelings of isolation and disconnectedness, and lead to more effective learning outcomes. We investigated what behavioral and demographic factors contributed to students’ perception of classroom community in virtually conducted computing courses at a large public university in the Pacific Northwest region of the United States. We deployed a survey composed of self-reported demographic and behavior questions, and items from a previously validated instrument on classroom community for online courses. The questions were coded into 53 features used to predict a student’s sense of classroom community. We used three machine learning models to do this prediction: a random forest, a gradient boosting machine, and a neural network. These models were validated and then analyzed with SHapley Additive exPlanations (SHAP), a relatively new modelagnostic technique which uses principles from game theory to explain how data features influence a complex model’s output. The most impactful prediction features for an individual student’s self-assessed sense of classroom community across all three models were whether the student was an international student (negative impact), and participation behavior, including whether a student spoke, used the chat feature, interacted in breakout rooms, or turned on a webcam (all positive impact).