多视图学习的基础理论和方法研究

Research on the Fundamental Theory and Methods of Multi-View Learning

随着信息获取技术的不断发展,现实中往往能收集到一个事物多个方面的数据,一般称为多视图/多模态数据。多视图学习有效地协同利用了多视图信息,可以更好地揭示数据中存在的潜在模式,在许多实际应用中带来了显著成效。本文聚焦于面向复杂多视图数据的表示、分类、建模与理论,阐述了以下两个方面:一是如何有效利用不同视图之间的共性(相关性或一致性)和个性(独立性或互补性)进行建模;二是如何针对部分视图残缺的多视图数据进行建模。

With the rapid development of information acquisition technology,one always collects the target data from different aspects(usually denoted multi-view/multimodal data).Multi-view learning methods can effectively unify them so as to better uncover the intrinsic underlying data patterns,and are widely employed in a large set of real-word applications.This paper focuses on the representation learning,modeling and theory analysis of the complex multi-view data.Specifically,two bottleneck challenges are addressed:①how to effectively build a multi-view model by exploring the consensus(correlation or consistence)and view-specific(independence or complementarity)information among data views;②how to deal with the partially incomplete data.