Using co-training to empower active learning
Payam V. Azad, Yusuf Yaslan · 2017
Active Learning and co-training are cases of semi-supervised learning both are used when labeled data is scarce. Active learning attempts to improve learning model by querying over unlabeled data and the main challenge there, is to find the optimum instance query. And co-training tries to exploit two different feature sets to enlarge number of labeled data without any need to get external information. Several researches tried to couple these two methods and get best out of them and they achieve noteworthy results. But we have witnessed that using co-training and active learning in sequence architecture outperforms when they are working in parallel. Using them in sequence means we have used co-training techniques to just find the best queries for active learning, and not in learning process itself. We will demonstrate that it has better results than plain active learning and co-training and even current parallel architectures. For this work we have used different techniques to split data into two distinct datasets; we will also discuss about it alongside our query selection method.