Research And Implementation On Anchor Graph Based Multi-View Learning Algorithm

In multi-view data,all the instances are characterized by the data collected from multiple measurement methods.Combining the data from multiple views can often obtain more information than that from a single view.How to properly fuse multi-view features to improve algorithm performance has become a core issue for multi-view learning.Graph-based multi-view learning has become one of the most popular methods in multiview learning due to its ability to extract nonlinear structures in multi-view data.However,most of them exhibit high computational complexity since they have to decomposing large-scale graph adjacency matrices in optimization.To overcome this drawback,researchers introduced the anchor graph to improved the speed of the graphbased methods.The key idea of the anchor graph strategy is to introduce a group of representative samples named anchors,and construct an anchor graph between the anchor set and the sample set,where the links between them indirectly represent the relationship among samples.Then utilizing the sparse and block characteristics of the anchor graph adjacency matrix to quickly complete large-scale matrix decomposition and other operations,accelerating the graph-based learning task.It can be seen that the construction and fusion of anchor graphs is the core task of anchor graph-based multi-view learning.How to construct anchor graphs based on raw data to mine the latent structures in multiview data,and how to efficiently fuse multi-view anchor graphs are two issues that are worth to study.Aim to the problems above,this paper studies the multi-view learning algorithm based on anchor graph,and the main results are as follows.For the problem of how to efficiently fuse multi-view anchor graphs,we propose a new multi-view anchor graph fusion method named Structure Diversity-induced Anchor Graph Fusion(SDAGF),and apply it to the multi-view clustering problem.In existing anchor graph based methods,the anchor generation method is severely limited by the multi-view anchor graph fusion strategy.In order to obtain valid fusion results,the existing graph fusion strategies require the vertices in the graph represent the same instance(i.e.anchors and samples),which restricts the anchor generation to be same on different views,degenerating the representation ability of corresponding fused consensus graph,and losing the structural diversity of each view.merge the anchors and edges of all the view-specific anchor graphs into a single graph for the structural optimal graph learning.Benefiting from the structural fusion strategy,the anchor generation of each view is not forced to be same,which greatly improves the representation capability of the target structural optimal graph,since the anchors of each view capture the diverse structure of different views.By leveraging the potential structural consistency among each anchor graph,a connectivity constraint is imposed on the target graph to indicate clusters directly without any post-processing such as 6)-means in classical spectral clustering.Substantial experiments on real-world datasets are conducted to verify the superiority of the proposed method,as compared with the state-of-the-arts over the clustering performance and time expenditure.To address the problem of how to construct anchor graphs based on raw data to mine the latent structure in multi-view data,we propose a method named Bipartite graph derived Embedding for Multi-View Learning(BEMVL),and apply it to into the multiview multi-label classification problem.Existing anchor graph-based fusion algorithms usually focus on the local structure between anchors and samples while ignoring the global structure.BEMVL adaptively learns anchors to construct anchor graphs while constraining anchor graph to contain global structure through distance-preserving mapping.The global structure of the original graph improving the quality of the learned anchor graph and improving the classification ability of the multi-view multi-label learning method.Further,we design a concise but efficient feature collinearity-induced feature selection method to learn compact multi-label classifiers.The resulting objective function is optimized in an alternating optimization fashion.Experimental results on different multi-label image datasets verify the efficiency and effectiveness of the proposed method.