Multi-View Hierarchical Semi-supervised Learning by Optimal Assignment of Sets of Labels to Instances

In multiclass semi-supervised learning, sometimes the information about datapoints is present in multiple views. In this paper we propose an optimization based method to tackle semi-supervised learning in the presence of multiple views. Our techniques make use of mixed integer linear programming formulations along with the EM framework to find consistent class assignments given the scores in each data view. We compare our techniques against existing baselines, including a cotrain variant for K-Means, on a number of multi-view datasets. Our proposed techniques give state-of-the-art performance in terms of F1 score, outperforming a well-studied SSL method based on co-training. Further, we show that our techniques can be easily extended to multi-view learning in the presence of hierarchical class constraints. These extensions improve the macro-averaged F1 score on a hierarchical multi-view dataset.