Affiliations: [a] School of Information Technology and Engineering, University of Ottawa, Ottawa, ON, Canada | [b] Institute for Computer Science, Polish Academy of Sciences, Warsaw, Poland
Correspondence:
[*]
Corresponding author: Jin Huang, School of Information Technology and Engineering, University of Ottawa, Ottawa, ON, Canada K1N 6N5. E-mail: jhuang33@site.uottawa.ca.
Abstract: Semi-supervised learning algorithms are widely used to build strong learning models when there are not enough labeled instances. Some semi-supervised learning algorithms, including co-training and co-EM, use multiple views to build learning models. Past research has shown that multi-view learning usually shows advantages over learning with a single view. However, conditions such as independence of the views, which is often hard to achieve, makes successful application of such methods in real world problems difficult. We would like to know if the performance of a multi-view semi-supervised learning method can be improved even though the available views are not necessarily independent. In this paper, we propose a simple sampling method, agreement-based sampling, as one way of improving the classification performance of multi-view semi-supervised learning algorithms. We apply agreement-based sampling to three major multi-view semi-supervised learning algorithms: co-training, co-EM and multi-view semi-supervised ensemble. The experiments with real-world datasets show that our sampling method can indeed significantly improve the performance of these three algorithms. We also investigated the relation between the quality of the expanded labeled set generated by the semi-supervised learning algorithm and the classification error rate of the learned model.
