Multi-view subspace clustering with intactness-aware similarity

Highlights：

• We construct the similarity based on the assumption that the target similarity has maximum dependence with its corresponding intact space, which can be measured by the Hilbert–Schmidt Independence Criterion (HSIC). Moreover, a new explanation (local connectivity) on such a similarity has been provided, that is, the learned intactness-aware similarity will have a larger value if their data points in the intact space have a small l1 distance.

• We demonstrate the efficacy and the superior performance of our proposed framework over the state-of-the-arts by conducting experimental results on six benchmark datasets.

摘要

•We propose a novel multi-view subspace clustering model, namely Multi-view Subspace Clustering with Intactness-Aware Similarity (MSC_IAS), that constructs the similarity based on the intact space learning technique, and unifies them into one framework. Moreover, two separated steps are conducted to validate the superiority of such the joint optimization.•We construct the similarity based on the assumption that the target similarity has maximum dependence with its corresponding intact space, which can be measured by the Hilbert–Schmidt Independence Criterion (HSIC). Moreover, a new explanation (local connectivity) on such a similarity has been provided, that is, the learned intactness-aware similarity will have a larger value if their data points in the intact space have a small l1 distance.•We demonstrate the efficacy and the superior performance of our proposed framework over the state-of-the-arts by conducting experimental results on six benchmark datasets.