英文摘要 | In the big data era, manners for data collection are more diverse, and forms of data representation are more various. Thus, the object of observation usually could be represented by different features, which is named as multi-view data by academic community. For example, image could be described by both global feature and local feature. The former reflects the overall character of the image, while the latter is based on significant areas or key points of the image. In order to exploit the rich embedded information among the multiple views, multiview learning emerges at the right moment and becomes a hot topic in machine learning field. Subspace learning aims to map the data from a high-dimensional original feature space into a low-dimensional subspace, and maintain some certain statistic characters at the same time, which could alleviate curse of dimensionality effectively. However, many classic subspace learning methods, such as topic model, matrix factorization, always ignore the internal attributes among multiple views, and cannot handle the multi-view data well. This paper focuses on subspace-based multi-view learning. Different from classic subspace learning, which transforms data from one high-dimensional original space into one low-dimensional subspace, subspace-based multi-view learning tries to discover a unified low-dimensional subspace from multiple highdimensional original spaces and obtain a unified feature representation which embeds multi-view information. Subspace-based multi-view learning, on one hand keeps the advantage of subspace learning that it could alleviate curse of dimensionality effectively, on the other hand, it could make full use of the multiview data. This paper firstly discusses two kinds of multi-view internal attributes, i.e., consistency and complementarity, which ensure the effectiveness of multi-view learning. And then, under the unsupervised and semi-supervised conditions, this paper exploits consistency and complementarity, and proposes some effective subspace-based multi-view learning methods. In addition, this paper also applies the thoughts of multi-view learning into a type of quasi-multi-view learning problem. Main contributions of this paper are summarized as follows: 1. This paper proposes two kinds of unsupervised multi-view learning methods in the frame of probabilistic latent semantic analysis with co-regularization idea, i.e., topic-based co-regularied probabilistic latent semantic analysis and pair-bas... |
修改评论