Deep Embedded Subspace Clustering with Hard-Sample Mining

In recent years, subspace clustering has received increasing attention for its ability to accurately discover the underlying subspace
in high-dimensional data. Among them, end-to-end subspace clustering methods compute cluster assignments by mapping points to subspaces.
However, such methods ignore hard samples when computing the clustering assignment, i.e., low-value samples in the correct clustering assignment and high-value samples in the incorrect clustering assignment.To address this problem, in contrast to the previous instance-level hard samples, we mine hard samples at the subspace cluster-level. We first construct a deep embedded subspace clustering framework as the clustering target and learn subspace bases in iterations to obtain clustering assignments. Secondly, we utilize pseudo-supervised information and clustering assignments to mine hard samples at subspace cluster-level.Finally, a weight modulation strategy is proposed to dynamically focus the hard samples and obtain more accurate subspace clustering assignments. Through extensive experiments, we show that our method outperforms state-of-the-art subspace clustering algorithms on four benchmark datasets.