Instances Based Multi-view Classification Model

The goal of multi-view learning is to improve the performance of classifiers by using multi-view data collected by different measuring techniques and methods,because data from different views have the characteristics of consistency and complementarity.In recent years,a variety of kinds of multi-view algorithms have been proposed,especially the multi-view algorithm based on SVM,which has a solid theoretical basis,has been widely studied and applied.However,existing algorithms are restricted by the assumption that two views are equally important in the learning process and the decision functions of different views in the final fusion step must also be averaged.Actually,different data instances have different characteristics.Therefore some instances can be described by one view better,while others are represented better in another feature space.Various instances have individual adaptability for different views,and the information of different views has individual description accuracy for various instances.The existing multi-view algorithms ignore the feature of multi-view instances.In this paper,we propose the concept of view vector of an instance,which describes the importance of different views for a specific instance and gives the weight of each instance with respect to different views.The view vector is obtained by training and classifying the data with multiple classifiers selected in advance and calculating the results statistically.Next,we combine the view vector and SVM model,and apply it to the training and testing phases of the model in order to build a new multi-view SVM model based on the instance.Compared with the existing multi-view classification model,the new model focuses on the characteristics of each data instance,rather than generalizing them.The application of view vector makes the model fully consider the importance of the information contained in individual features space of different training instances.The more important the information ofview is,the greater role it plays in the process of training the model.On the contrary,unimportant information of view has little effect on model training.The instance-based multi view model also combines the view vector of the instance with the final result fusion process in the model classification stage so that the larger the value of the vector component is,the greater the weight of the classification result is,because such a view is more accurate and important for the instance.Then we apply the view vector to the MVNPSVM algorithm,making it act on the fusion stage of the classification results,and it makes the original MVNPSVM algorithm pay more attention to the characteristics of the data instance itself,and further improve the classification accuracy of the algorithm.For the instance-based multi-view SVM model and the improved MVNPSVM algorithm,we have used a large number of multi-view data sets to fully verify the theory,and we also compared the new algorithm with other excellent multi-view models.The experimental results show the superiority of the new classification model.