Prior Indicator Guided Anchor Learning for Multi-View Subspace Clustering

Abstract

Despite multi-view subspace clustering (MVSC) has the widespread utilization in practical scenarios, the majority of existing methods face challenges when confronted with large-scale data. To address this issue, the anchor strategies have been proposed, but they mainly face the following problems: (1) most methods use heuristic anchor sampling results in weak discrimination of anchors, as it separates anchor sampling and graph construction; (2) the distribution of anchors within the cluster is uneven, and have low goodness-of-fit between anchors and raw data. To tackle the aforementioned issues effectively, we propose a method named Prior Indicator Guided Anchor Learning for Multi-view Subspace Clustering (PIAL). Specifically, PIAL proposes a unified framework, which can adaptively learn anchors and bipartite graph. Most importantly, PIAL introduces the prior indicator to constrain the bipartite graph learning. In this way, these two frameworks synergistically promote each other to acquire a high-quality anchor graph, such that more flexible and discriminant anchors can be obtained. Moreover, PIAL allows large-scale data to be processed in linear time, which is beneficial for large tasks. Experiments are carried out on several real-world benchmarks and compared with some state-of-the-art methods to verify the comparability and availability of PIAL.