稀疏表示一致性引导的多视图降维算法

Sparse Representation Consensus for Multi-view Dimensionality Reduction

现有基于图的多视图降维方法大多将构图和降维两个过程独立执行,因此构图的质量直接决定着降维的效果,然而构图是一个开放性的问题.为了缓解上述困难,提出了一种稀疏表示一致性引导的多视图降维算法(MDR_SRC).首先,通过使不同视图下的样本保持公共的稀疏表示,挖掘了视图之间的一致性关系;其次,根据样本对稀疏表示系数的差异性指导构图,利用构建的图指导降维;然后将基于稀疏表示的构图与基于图的降维整合为一个优化问题,使构图与降维过程相互指导,从而实现图质量的动态提升;最后,设计了一种迭代地交替策略求解该优化问题.在4个公开数据集上的实验结果表明文中所提的方法优于现有的代表性多视图降维方法.

Most of the existing graph-based multi-view dimensionality reduction methods carry out the processes of graph construction and dimensionality reduction independently,so the quality of graph determines the performance of dimensionality reduction directly.However,how to construct a high-quality graph is an open problem.In order to alleviate the above difficulties,a Sparse Representation Consensus for Multi-view Dimensionality Reduction(MDR_SRC)method is proposed.First of all,the consensus relationship between multiple views is mined by keeping the common sparse representation of samples under different views.Secondly,the difference be-tween the sparse representation coefficients of sample pairs is used to guide the graph construction,and the constructed graph is used to guide the dimensionality reduction.After then,the graph construction based on the sparse representation and the dimensionality reduc-tion based on the graph are integrated into an optimization problem,so as to realize the mutual guidance of graph construction and di-mensionality reduction.At the same time,the quality of graph is improved dynamically.Finally,the iteratively alternating strategy is designed to solve the optimization problem.The experimental results show that the proposed method is superior to the existing repre sentative multi-view dimensionality reduction methods on four public data sets.