Using Similarity Learning with SBERT to Optimize Teacher Report Embeddings for Academic Performance Prediction

Abstract

Student performance prediction continues to be a focus of research in educational data mining due to its many potential benefits. While teachers’ assessment reports are a crucial part of the educational process, they have not been commonly used in performance prediction. We propose a model that uses similarity learning as an embedding-enhancing technique. Results outperform earlier research with an average accuracy of 73% for detecting strong performance.