Distributed Sparse Subspace Clustering by K-Means Subspace Fusion

Sparse subspace clustering (SSC) achieves evidenced success in many areas but most related works focus on the centralized scenario. In many practical applications such as security surveillance, several distributed agents jointly collect a large dataset, while each one is however just able to observe a subset through which a local inference is to be made. This paper proposes a collaborative distributed SSC scheme targeted for this situation. In the proposed scheme, each agent first conducts a fast local data clustering, and transmits basis matrices of the estimated subspaces to the data center, which then conducts subspace information fusion using a k-means type method; the aggregated subspace information is fed back to local agents to update their data partitions. Computer simulations using real human face data are used to illustrate the effectiveness of the proposed method.