Classifier Design And Weight Optimization Methods Based On Multiple Views

This paper researches new algorithms to improve the learning performance of Multi-View models. These methods can not only optimize the weight distributions effectively, but also design the classifier models based on Multi-View efficiently. The main work of this paper is as follows:1. A new optimal learning algorithm, called RMultiV-MHKS algorithm, is proposed. This algorithm combines the surface response algorithm (RST) with the traditional MultiV-MHKS. The key idea of the method is to use weight variables and classification errors together to form the optimal response surface by limited iteration steps with the Newton method. Accordingly, the more reasonable weights of Multi-View Machine Learning models are obtained and thus their classification accuracies are improved effectively.2. The existing surface response algorithm (RST) is introduced to the classifier MultiK-MHKS, and thus a new RMEK-MHKS algorithm is proposed, in order to find the optimal weight values with kernels and improve the accuracies of Multi-Kernel Machine Learning models. A lot of experimental results prove the advantages and versatility of this method.3. By fully considering the impact of samples on the resulting classification models, we introduce a local "control" function (Modal Gating) into the MultiK-MHKS design and build a new Multi-Kernel Machine Learning model named MLEKL. The iteration cycle process of finding the best classification rate is mainly limited to the internal optimization of local "control" function, so we can use the sample information to improve the classifi-cation accuracy of the MLEKL model, and effectively imrpove the training efficiency of the algorithm.The above-mentioned methods optimized the weight distribution on the traditional Multi-View model to some extent and got the comprehensive optimal values of weights. The response surface method can fully optimize the weight values of views, but has higher time complexity. Contrastively, local "control" function can make the best of information to ensure a Multi-View or Multi-Kernel classifier classification accuracy, and at the same time, it can effectively limit the time complexity. Finally, whether through experimental verification on common data set of Multi-View, or graphical analysis on the classification of artificial data sets can prove the superiority of the proposed methods.