Subspace clustering via seeking neighbors with minimum reconstruction error

Abstract

Subspace clustering refers to segmenting data in accordance with the underlying subspaces. To this end, the state-of-the-art methods commonly obtain representations of samples on a dictionary to measure similarities among samples. However, most these subspace clustering methods are still very sensitive to noise in the data. To relieve the impact of the noise, we propose an iterative method by successively seeking k samples (k Sparse Representation Neighbors, k SRNs) with the highest similarity to a query. In each round, based on a fast constructed dictionary, a sparse representation is obtained using Lasso regression. Afterwards, we introduce a similarity measurement relevant with reconstruction error. By that means, the sample with minimum reconstruction error is selected as a SRN which is utilized to update the dictionary in the next round. Since each SRN has the minimum reconstruction error, the proposed method is robust to the noise in samples. Experimental results on the commonly used benchmark datasets show that the proposed method outperforms existing methods.