Zero-shot Learning for Predicting Unseen Student Activities in Educational Platforms

Abstract

Zero-shot learning (ZSL) has emerged as a groundbreaking approach in machine learning, enabling models to classify unseen categories by leveraging the relationships between known and unknown categories through semantic knowledge. Within educational platforms, accurately predicting and understanding student activities is vital for personalizing learning, tracking progress, and improving adaptive learning systems. However, the wide diversity of potential student activities makes collecting labeled data for every scenario unfeasible, posing a significant limitation to traditional supervised learning methods. This study addresses this limitation by introducing a ZSL framework specifically designed to predict unseen student activities. The proposed framework leverages semantic embeddings, such as word2vec and BERT, to establish meaningful connections between known and unknown activities, enabling accurate predictions without labeled examples. The framework is thoroughly evaluated using real-world datasets of student interaction logs, with performance assessed across metrics such as accuracy, precision, recall, and F1-score. The results highlight the framework's ability to deliver strong predictive performance while providing valuable insights into the relationships between activity categories. By bridging the gap between labeled and unlabeled data, this research showcases the transformative potential of ZSL in advancing educational platforms. It demonstrates how ZSL can enhance adaptive learning systems, foster student engagement, and equip educators withactionable insights, driving the development of smarter, more personalized educational technologies.