Two novel deep multi-view support vector machines for multiclass classification

Abstract

Multi-view classification methods have better generalization performance compared to the single-view classification methods due to the consistency information from multiple views. In recent years, the combination of support vector machine (SVM) and multi-view learning has been widely studied. To improve the robustness of multi-view classification methods, emphasis has shifted to the integration of multi-view classification approaches with fully-connected and convolutional neural networks. A classical deep two-view classification method named deep SVM-2K is a combination of support vector machine with two stage kernel canonical correlation analysis (SVM-2K) and deep learning. However, limitations of deep SVM-2K are that it can not cope with multi-view classification and multiclass classification problems. To address these issues, we propose two novel deep multi-view models named deep multi-view support vector machine (DMVSVM) for multiclass classification. DMVSVM uses the learned features by auto-encoder (AE) or deep neural network (DNN) to train the SVM classifier for each view. The two models then impose some constraints to make the output of the multi-view SVM classifiers as consistent as possible, which used to exploring intrinsic relations. Experiments performed on different real-word datasets show the effectiveness of our proposed approaches.