Multiview Classification with Missing-Views Through Adversarial Representation and Inductive Transfer Learning

Abstract

The importance of Multiview learning in classification task cannot be further stressed. Over the years, there is an increasing interest among researchers to enhance Multiview classification models. However, very little attention has been paid to handling the missing-view(s) problem in existing systems. Missing-view can be caused by various reasons. It can be a technical fault during data measurement which can occur in real-world applications. Handling such occurrences is paramount to determine the success of the classification model on real-world data. In this paper, a Multiview deep classification model is proposed. The model implements an adversarial network that learns a latent space of semantic knowledge of the Multiview data. Further, the model used an inductive transfer style to train a neural classifier network from the learned latent space. Therefore, understanding the semantics is required to recreate any corresponding alternate view from a given view. The proposed model is trained and evaluated on MNIST and Noisy MNIST datasets. Its performance was compared with existing systems in terms of accuracy and normalized mutual information (NMI). The proposed model achieved better performance. The model is evaluated across different missing rates which further demonstrated that it is capable of handling missing-view(s) effectively. © 2022, Springer Nature Switzerland AG.