Spectral Clustering by Subspace Randomization and Graph Fusion for High-Dimensional Data

摘要

Subspace clustering has been gaining increasing attention in recent years due to its promising ability in dealing with high-dimensional data. However, most of the existing subspace clustering methods tend to only exploit the subspace information to construct a single affinity graph (typically for spectral clustering), which often lack the ability to go beyond a single graph to explore multiple graphs built in various sub-spaces in high-dimensional space. To address this, this paper presents a new spectral clustering approach based on subspace randomization and graph fusion (SC-SRGF) for high-dimensional data. In particular, a set of random subspaces are first generated by performing random sampling on the original feature space. Then, multiple K-nearest neighbor (K-NN) affinity graphs are constructed to capture the local structures in the generated subspaces. To fuse the multiple affinity graphs from multiple subspaces, an iterative similarity network fusion scheme is utilized to achieve a unified graph for the final spectral clustering. Experiments on twelve real-world high-dimensional datasets demonstrate the superiority of the proposed approach.