基于自表达的多视角子空间聚类方法研究

发布时间：2018-04-16 20:08

本文选题：多视角聚类 + 子空间聚类　；参考：《天津大学》2015年博士论文

【摘要】：多视角聚类要解决的关键问题是如何有效融合来自多个不同视角的信息,从而更加准确地发现数据潜在的结构,提高聚类效果。当前尽管存在一些基于多视角数据的学习方法,然而,大多数多视角学习方法主要聚焦于监督或者半监督学习,依赖于标注数据指导学习过程。相对于标注数据,现实中的无标注数据非常丰富。随着计算机在计算、存储能力方面的极大提高,有效利用大规模、无标注的数据越来越重要。然而,由于缺少标注数据的指导,无监督或者弱监督条件下的数据分析更具挑战性。本文聚焦于无监督或弱监督条件下多视角数据上的聚类方法,研究如何有效融合多视角信息及先验约束提升聚类效果。具体来讲,主要包括以下三个方面的研究内容:1)增强互补性的多视视角聚类针对当前多视角聚类方法中,在构建各视角相似度矩阵时忽略了视角之间的互补性约束的问题,提出了多样性诱导的多视角子空间聚类方法。利用希尔伯特-斯密特独立性标准做为多样性度量,挖掘多视角子空间表达的互补性。2)高阶关联的多视视角聚类针对当前的多视角聚类方法仅利用成对的视角关联而忽略了多视角之间的高阶关联的问题,提出了低秩张量约束的多视角子空间聚类,通过引入低秩张量约束来挖掘和利用多视角之间的互补性。3)融合约束的多视视角聚类针对聚类过程中存在的先验信息,提出了约束的多视角聚类方法,聚焦于同时有效利用先验约束和多视角一致性提升聚类效果。并以视频人脸聚类做为应用范例,提出了约束的多视角视频人脸聚类框架。所提出的三种方法验证了挖掘视角之间的互补性、减少视角之间的冗余以及有效利用约束信息是提高聚类效果的有效手段。
[Abstract]:The key problem to be solved in multi-view clustering is how to effectively integrate the information from different perspectives so as to find the potential structure of data more accurately and improve the clustering effect.Although there are some learning methods based on multi-view data, most of them focus on supervised or semi-supervised learning and rely on annotated data to guide the learning process.Compared with annotated data, there are plenty of unannotated data in reality.With the improvement of computer computing and storage capacity, it is more and more important to utilize large scale and unlabeled data effectively.However, data analysis under unsupervised or weakly supervised conditions is more challenging due to the lack of guidance of annotated data.This paper focuses on the clustering method of multi-view data under unsupervised or weakly supervised condition, and studies how to effectively fuse multi-view information and priori constraints to improve the clustering effect.Specifically, this paper mainly includes the following three aspects: 1) Multi-view clustering, which enhances complementarity, ignores the complementary constraints in constructing the similarity matrix of different visual angles in the current multi-view clustering method.A multi-view subspace clustering method for diversity induction is proposed.Using the Hilbert-Smit independence criterion as a measure of diversity,Mining complementarity of Multi-view subspace expression. 2) High-order relevance Multi-view clustering aiming at the problem of high-order correlation between multi-view angles, the current multi-view clustering method only uses pairwise visual angle association, but ignores the problem of high-order correlation between multi-view angles.In this paper, a multi-view subspace clustering with low rank Zhang Liang constraint is proposed. By introducing the lower-rank Zhang Liang constraint to mine and utilize the complementary .3between multi-view constraints, the multi-view view clustering is aimed at the prior information existing in the clustering process.A multi-view clustering method with constraints is proposed, which focuses on using both priori constraints and multi-view consistency to improve clustering performance.Taking video face clustering as an application example, a constrained multi-view video face clustering framework is proposed.The three methods proposed in this paper verify the complementarity between the mining angles, reduce the redundancy between the angles and effectively utilize the constraint information to improve the clustering effect.
【学位授予单位】：天津大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP311.13

【相似文献】