Multi-view Learning With The Universum

In recent years, multi-view learning is a hot field of machine learning research directions. Multi-view learning learns how to use two separate things or irrelevant perspective with specific training approach to learning. The traditional Multi-view Learning (MVL) learns how to pro-cess patterns with multiple information sources. In generalization the MVL is proven to have a significant advantage over the usual Single-view Learning (S VL). But in most real-world cases, we only have single source patterns to deal with, where the existing MVL is unable to be direct-ly applied. In order to solve this problem, people developed a novel MVL technique for single source patterns through reshaping the original vector representation of single source patterns into multiple matrix representations. Doing so can effectively bring an improving classification accuracy. In this paper, we aim to generalize the above MVL through taking advantage of the Universum examples which do not belong to either class of the classification problem. Through these Universum samples, we can get a prior domain knowledge of the whole data distribution. The main work of this paper is as follows:· In implementation, we incorporate our previous MVL instance named MultiV-MHKS with the available Universum examples, and thus get a more flexible multi-view classification machine with the Universum called UMultiV-MHKS for short. The new method UMultiV-MHKS is based on the MultiV-MHKS and add some Universum samples, so our approach have a Universum regularization term than MultiV-MHKS. In order to demonstrate the effectiveness of our proposed algorithm, we will do some experiments with our algorithm and compare with several other similar algorithms in classification accuracy and training time. Then, we will have a further discussion on our proposed algorithm with three aspects, such as the selection of the regularization parameter values, the selection of the scale of the Universum sample and the analysis of the convergence. Finally, through the analysis of the complexity, We prove that UMultiV-MHKS get a more rigorous error bounds than MultiV-MHKS.· In order to demonstrate whether the selection method of the Universum samples have impact on experiments, we also discuss the selection the selection methods of the Universum samples separately. We selected two existing selection methods:Method based on geometric models and the nearest neighbor selection methods. Then, we compare these two methods with our methods Umean which is used in our experiments. Through analysis the result of the experiment, we can conclude that different selection methods of the Universum samples have little impact on our experiment.