Machine Learning | Safe Semi-Supervised Learning

Semi-supervised learning (SSL) concerns the problem of how to improve learning performance via the usage of a small amount of labeled data and a large amount of unlabeled data. Many SSL methods have been developed, e.g., generative model, graph-based method, disagreement-based method, and semi-supervised SVMs. Despite the success of SSL, however, a considerable amount of empirical studies reveal that SSL with the exploitation of unlabeled data might even deteriorate learning performance. It is highly desirable to study a safe SSL scheme that on one side could often improve performance, on the other side will not hurt performance, since the users of SSL won't expect that SSL with the usage of more data performs worse than certain direct supervised learning with only labeled data. #SafeSSL #DataScience #MachineLearning
======================================================================