Parallel interacting multiview learning: An application to prediction of protein sub-nuclear location

In some machine learning problems, the dataset has multiple views which may be obtained using different sensors or applying different sampling techniques. These views may have sufficient or partial information about the target concept. In this paper, a method that we called parallel interacting multiview learning (PIML) is proposed in which the views interact during the training process using the predictions of each other together with their original features. This way, the views are expected to strengthen the prediction accuracies of the other views feeding their predictions to the others even during the training process. This technique avoids the way of simply merging features of all views and reaches higher accuracy than its counterparts that do not interact during learning but only combine their predictions after the learning process. PIML is demonstrated on a real bioinformatics dataset for predicting protein sub-nuclear locations.