基于Universum数据的多视角学习算法

Multi-View Learning Algorithm Based on Universum Data

摘要: 多视角学习是以不同方法获得的特征集表示的数据中学习的问题，其中双视角学习是一种仅由双视角数据组成的多视角学习。由于多视角学习可能会忽略一些多视角数据的原始信息，这些数据之间存在着内在的联系和不同视角之间的差异。因此，为了解决多视角数据之间存在的问题，我们引入了既不属于正类又不属于负类的无标签数据Universum数据。本文提出了一种基于Universum数据的多视角学习算法，将Universum数据和多视角学习结合到一个目标模型中，其中Universum数据被认为是该模型的先验知识。为了解决提出的算法模型，我们推导了该算法模型的对偶问题并得到了预测分类器。最后，通过大量的实验对该方法的性能进行了研究，结果表明，所提算法的性能优于传统的方法。

Abstract: Multi-View Learning (MVL) focuses on the problem of learning from the data represented by feature sets obtained from different approaches, in which two-view learning is a kind of MVL which only consists of two-view data. Since multi-view learning may ignore the original information of some multi-view data, there are inherent connections between these data and the differences between different perspectives. Therefore, in order to solve the problems between multi-view data, we introduce the unlabeled data Universum data that neither belongs to the positive category nor the negative category. This paper proposes a multi-perspective learning algorithm based on Universum data, which combines Universum data and multi-perspective learning into a target model, in which Universum data is considered as the prior knowledge of the model. In order to solve the proposed algorithm model, we derive the dual problem of the algorithm model and get the predictive classifier. Finally, the performance of the method is studied through a large number of experiments, and the results show that the performance of the proposed algorithm is better than that of the traditional method.