Keeping Students in their Zone of Proximal Development: A Machine Learning Approach for Intelligent Tutoring Systems

Students benefit from working on content that is challenging without being too difficult. However, even in adaptive learning systems, ensuring that the content meets these criteria for all students can be complex. To complement placement tests and knowledge tracing in mastery learning activities, we use students' performance to classify whether they are working outside of their Zone of Proximal Development (ZPD). We demonstrate that using estimates of student achievement and lesson difficulty, we can predict whether students are working outside of their ZPD in upcoming lessons with the goal of moving them to content that aligns with the ZPD. We train and compare different machine learning models on a dataset from a math-focused intelligent tutoring system, and find that the XGBoost model achieves the best performance predicting whether students are working above their ZPD (AUC of .83) as well as below their ZPD (AUC of .79). Our models predict students working outside of their ZPD with at least 80% precision using data from as few as one lesson. This work can be used to enable more flexible adaptivity for Intelligent Tutoring Systems.