A semi-supervised ensemble learning algorithm

The success of ensemble learning usually depends on available labeled data. We present a new and more general co-training style framework, ensemble co-training, to combine ensemble learning with semi-supervised learning. A new classifier is generated using both labeled data and the most confident newly-predicted data, and added into the ensemble at each round. Finally, the most credible classifiers are selected for the final prediction. Furthermore, we provide a theoretical analysis for the learnable ability of co-training style algorithms in the presence of both classification noise and distribution noise. We demonstrate our algorithm on six text datasets, and the results show that the ensemble co-training performs better than other state-of-the-art algorithms in practice.