大数据多视图子空间非监督机器学习理论与方法

With the arrival of the era of big data, the demand of intelligent processing method for high-dimensional heterogeneous data has become increasingly prominent, and then the unsupervised multi-view learning is considered as the key to knowledge acquisition. Existing unsupervised multi-view learning approaches lack a comprehensive theory to guide the establishment of the relationship among different views, as well as can’t provide the robustness and scalability in practical applications. Accordingly, it is urgent to carry out the study on essential theory and method for multi-view subspace clustering. This project aims to propose a novel multi-view subspace representation model via high-order tensor factorization, so as to make a breakthrough on critical problems such as multi-view feature learning, handling on view data missing, and algorithm parallelization. Firstly, a novel theory, which is built upon the tensor low rank regularization, is proposed to explore the complementary and consistency among different views, such that the high-order correlations underlying multi-view data can be captured. Secondly, study the following key techniques: multi-view subspace representation and affinity matrix simultaneous optimization, incomplete multi-view subspace clustering, and distributed parallel multi-view subspace clustering. Furthermore, a prototype system will be developed to validate the proposed theories and techniques. Through the aforementioned researches, a relatively comprehensive unsupervised learning framework will be established for large-scale multi-view data, which can support for the development of machine learning theory in the era of big data.

大数据时代海量多模态高维异构数据智能处理的需求日益凸显，非监督多视图机器学习已经成为知识获取的关键。当前多视图非监督学习方法缺乏视图间关联建模的核心理论指导，无法应对复杂条件下对算法鲁棒性和可扩展性的需求，迫切需要开展面向大数据的多视图子空间非监督机器学习理论与方法的研究。本项目旨在建立高阶张量的多视图子空间表示理论，突破子空间特征优化、视图样本残缺和并行可扩展性的瓶颈，从理论与模型、关键技术、典型实例验证三方面开展研究。研究内容：首先，研究高阶张量正则化的多视图互补和一致性建模和表示的核心理论；其次，研究多视图子空间特征与关联矩阵协同学习、视图残缺条件下的鲁棒多视图聚类以及高可扩展性的分布式并行多视图子空间聚类；最后，构建典型领域验证性实例，验证提出的理论及关键技术。通过上述研究，寻求较为完整的针对海量多视图特征的非监督学习框架，为大数据环境下的机器学习理论助力并提供关键支撑。

大数据时代海量多模态高维异构数据智能处理的需求日益凸显，非监督多视图机器学习已.经成为知识获取的关键。当前多视图非监督学习方法缺乏视图间关联建模的核心理论指导，无法应对复杂条件下对算法鲁棒性和可扩展性的需求，迫切需要开展面向大数据的多视图子空间非监督机器学习理论与方法的研究。本项目旨在建立高阶张量的多视图子空间表示理论，突破非线性子空间特征优化和高可扩展性多视图学习瓶颈，从理论与模型、关键技术、典型实例验证三方面开展研究并且获得重要进展。首先，研究高阶张量正则化的多视图互补和一致性建模和表示的核心理论，将传统的多视图聚类方法的时间复杂度降低了一个数量级同时获得了明显超越当前方法的聚类精度；其次，提出了两种非线性多视图子空间建模方法，有效应对高维特征分布不符合线性子空间的假设，同时提出了端到端的高可扩展性的深度多视图子空间聚类方法，较传统多视图非监督学习方法在数据处理规模上大幅提高并且获得了当前最高的聚类精度；最后，构建大规模图像检索和行人重识别领域验证性实例，验证提出的理论及关键技术。通过上述研究，寻求较为完整的针对海量多视图特征的非监督学习框架，为大数据环境下的机器学习理论助力并提供关键支撑。