QA-Knowledge Attention for Exam Performance Prediction

In actual university education, students’ performance prediction is important for assessing their mastery of specific knowledge areas and providing feedback. To address the limitations of existing grade or at-risk predictions that fail to take students’ knowledge mastery into account, we propose a model, question&answer knowledge attention for exam performance prediction (QAKAP) and focus on predicting students’ performance on each final exam question (binary prediction). This model consists of three key modules: a Topic-Aware Attention Module, a Knowledge Mastery Estimation Module, and a Prediction Module. By integrating natural language processing (NLP) methods and the attention mechanism, our model can capture the relevance of questions from question text data and then generate the student embeddings to predict student performance on each final exam question. The primary contribution of this work is the development of a general prediction model for real university education, applicable to students’ exam performance predictions. We collected a dataset from 494 students across four undergraduate courses, including the question text data and students’ performance on these exam questions. The results show that QAKAP outperforms other machine-learning methods, demonstrating its effectiveness in predicting students’ exam performance.