Semi-supervised Learning with Limited Labeled Data: Examining semi-supervised learning methods to leverage both labeled and unlabeled data for model training

Abstract

Semi-supervised learning (SSL) offers a compelling approach to training machine learning models when labeled data is scarce and expensive to obtain. This paper explores the effectiveness of SSL methods in scenarios where labeled data is limited, focusing on techniques that leverage both labeled and unlabeled data. We review prominent SSL algorithms and their applications, discussing their advantages and limitations. Additionally, we propose a novel SSL algorithm tailored for scenarios with severely limited labeled data. Through experiments on benchmark datasets, we demonstrate the efficacy of our approach compared to existing methods. Our findings highlight the potential of SSL in practical settings with limited labeled data, opening avenues for further research in this area.