A Large Margin Method for Semi-Supervised LearningAuthors: Xiaotong Shen, Junhui Wang and Wei Pan
Semi-supervised learning involves a large amount of unlabeled data with only a small number of labeled data. This imposes a great challenge in that the class probability given input can not be well estimated through labeled data alone. In this talk, I will present a large margin semi-supervised method based on an efficient margin loss for unlabeled data. This loss seeks extracting the maximum amount of information from unlabeled data for estimating the Bayes rule, when some prior knowledge is available for the class probability. It permits an integration of labeled and unlabeled data with regard to the Bayes rule. For implementation, an iterative scheme is developed. Finally, the method's generalization performance is investigated in addition to several numerical examples.
The work is done jointly with Junhui Wang (Columbia University) and Wei Pan (University of Minnesota).