Scholarlink

The predictive performance of learning models largely depends on the quality of the datasets used in Educational Data Mining (EDM). However, student performance datasets often vary in feature definitions and structures, which limits cross-dataset learning and predictive performance. The present paper presents a Large Language Model (LLM) based approach for harmonizing heterogeneous educational datasets to enable accurate student performance prediction. The proposed method leverages the LLM to align equivalent features across datasets, ensuring consistent representation of demographic and academic indicators. Applied to four datasets, three public and one private, the approach resulted in a unified dataset containing 31 harmonized features. LLM-based harmonization correctly retrieved matches in 0.8997 of the top 10 results, achieving a mean reciprocal rank of 0.3189 and outperforming traditional cosine and covariance similarity methods. When combined with a stacking ensemble model, it achieved 0.9081 accuracy, demonstrating the potential of LLM-guided harmonization to enhance predictive modeling in EDM.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Harnessing Large Language Models for Educational Dataset Harmonization for Student Performance Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Amani Khalifa,
  • Issam Rebai,
  • Fatma BenSaid,
  • Yessine Hadj Kacem

摘要

The predictive performance of learning models largely depends on the quality of the datasets used in Educational Data Mining (EDM). However, student performance datasets often vary in feature definitions and structures, which limits cross-dataset learning and predictive performance. The present paper presents a Large Language Model (LLM) based approach for harmonizing heterogeneous educational datasets to enable accurate student performance prediction. The proposed method leverages the LLM to align equivalent features across datasets, ensuring consistent representation of demographic and academic indicators. Applied to four datasets, three public and one private, the approach resulted in a unified dataset containing 31 harmonized features. LLM-based harmonization correctly retrieved matches in 0.8997 of the top 10 results, achieving a mean reciprocal rank of 0.3189 and outperforming traditional cosine and covariance similarity methods. When combined with a stacking ensemble model, it achieved 0.9081 accuracy, demonstrating the potential of LLM-guided harmonization to enhance predictive modeling in EDM.