Multi-view Representation Induced Kernel Ensemble Support Vector Machine

This paper proposes a multi-view representation kernel ensemble Support Vector Machine. Unlike the conventional multiple kernel learning techniques which utilizes a common similarity measure over the entire input space, with the aim of solely learning their models via linear combination of basis kernels in single Reproducing Kernel Hilbert Space (RKHS), our proposed model seeks to concurrently find multiple solutions in corresponding Reproducing Kernel Hilbert Spaces. To achieve this objective, we first derive our proposed model directly from the classical SVM model. Then, leveraging on the concept of multi- view data processing, we consider the original data as a multi - view data controlled by different sub models in our proposed model. The multi-view representations of the original data are subsequently transformed into ensemble kernel models where the linear classifiers are parameterized in multiple kernel spaces. This enables each model to co-optimize the learning of its optimal parameter via the minimization of a cumulative ensemble loss in multiple RKHSs. With this, there is an overall improvement in the accuracy of the classification task as well as the robustness of our proposed ensemble model. Since UCI machine learning data repository provides publicly available benchmark datasets, we evaluated our model by conducting experiments on several UCI classification and image datasets. The results of our proposed model were compared with other state-of-the-art MKL methods, such as SimpleMKL, EasyMKL, MRMKL, RMKL and PWMK. Among these MKL methods, our proposed method demonstrates better performances in the experiments conducted.