Scaled-Dot Product Attention for Early Detection of At-risk Students

Students’ performance prediction is essential for instructors to observe each student’s learning behavior to discover which students have become at-risk. Early prediction helps instructors to intervene in time and provide academic support to these students. However, instructors should grasp essential behavior points to survey students’ academic performance. In this study, we propose the Scaled-Dot Product Attention that can mine the relationship between the student’s learning behaviors and performance to find the essential features that directly affect students’ performance. In this study, we tested the early prediction performance of Scaled-Dot Product Attention with conventional algorithms. We then investigated essential lectures or features related to students’ learning activities. From the result, we found that Scaled-Dot Product Attention outperformed the conventional algorithms to identify at-risk students and found the important lectures and students’ actions.