Multi-view deep support vector machines based on discriminative contrastive loss

Abstract

Contrastive learning is a rapidly evolving direction due to its ability to learn outstanding discriminative representations. However, the two theoretically complementary models, contrastive learning and support vector machine (SVM), have never been integrated. In this paper, we propose two novel multi-view deep SVMs models based on the discriminative contrastive loss to solve the multi-view multi-class classification problem. Specifically, first, we impose the discriminative contrastive loss to learn the local structural information of each view. In addition, we propose a self-learning view-weight method to explore inter-view diversity information by assigning different view weights to each view, and explore cross-view consistency information by imposing similarity constraints on the disagreements generated by different view classifiers. Finally, two novel models take the $one-vs-rest$ method to solve the multi-class classification problem. In the optimization problem, the back propagation method is used to implement joint learning of the entire multi-view deep model. The experimental results validate the effectiveness of the models by comparing them with state-of-the-art algorithms.