Search for: All records

Teachers often use open-ended questions to promote students' deeper understanding of the content. These questions are particularly useful in K–12 mathematics education, as they provide richer insights into students' problem-solving processes compared to closed-ended questions. However, they are also challenging to implement in educational technologies as significant time and effort are required to qualitatively evaluate the quality of students' responses and provide timely feedback. In recent years, there has been growing interest in developing algorithms to automatically grade students' open responses and generate feedback. Yet, few studies have focused on augmenting teachers' perceptions and judgments when assessing students' responses and crafting appropriate feedback. Even fewer have aimed to build empirically grounded frameworks and offer a shared language across different stakeholders. In this paper, we propose a taxonomy of feedback using data mining methods to analyze teacher-authored feedback from an online mathematics learning platform. By incorporating qualitative codes from both teachers and researchers, we take a methodological approach that accounts for the varying interpretations across coders. Through a synergy of diverse perspectives and data mining methods, our data-driven taxonomy reflects the complexity of feedback content as it appears in authentic settings. We discuss how this taxonomy can support more generalizable methods for providing pedagogically meaningful feedback at scale. 

Knowledge Tracing models have been used to predict and understand student learning processes for over two decades, spanning multiple generations of student learners who have different relationships with the technologies used to provide them instruction and practice. Given that student experiences of education have changed dramatically in that time span, can we assume that the student learning process modeled by KT is stable over time? We investigate the robustness of four different KT models over five school years and find evidence of significant model decline that is more pronounced in the more sophisticated models. We then propose multiple avenues of future work to better predict and understand this phenomenon. In addition, to foster more longitudinal testing of novel KT architectures, we will be releasing student interaction data spanning those five years. 

The educational data mining community has extensively investigated affect detection in learning platforms, finding associations between affective states and a wide range of learning outcomes. Based on these insights, several studies have used affect detectors to create interventions tailored to respond to when students are bored, confused, or frustrated. However, these detector-based interventions have depended on detecting affect when it occurs and therefore inherently respond to affective states after they have begun. This might not always be soon enough to avoid a negative experience for the student. In this paper, we aim to predict students' affective states in advance. Within our approach, we attempt to determine the maximum prediction window where detector performance remains sufficiently high, documenting the decay in performance when this prediction horizon is increased. Our results indicate that it is possible to predict confusion, frustration, and boredom in advance with performance over chance for prediction horizons of 120, 40, and 50 seconds, respectively. These findings open the door to designing more timely interventions.