Jointly sparse neighborhood graph for multi-view manifold clustering

Abstract

View-specific neighborhoods commonly contain class-inconsistent neighbors in graph-based multi-view learning. A key problem is to handle class-inconsistent neighbors under each view. This paper employs jointly sparse learning to filter unreliable neighbors in the union of view-specific neighborhoods, via representing each entity in a weighted sum of its neighbors under each view. The proposed jointly sparse model can be easily solved by an ADMM method. The learned jointly sparse weights can be used to construct a similarity neighborhood graph, and the new graph can be further utilized for multi-view clustering and view-specific graph preconditioning. A fast algorithm for multi-view manifold clustering is then proposed, and two preconditioning approaches are discussed for improving conformability of view-specific graphs and eventually increasing the efficiency of graph-based algorithms for multi-view learning. Numerical experiments are reported, which provide good supports to the proposed methods.