Fast Multi-view Subspace Clustering with Balance Anchors Guidance

Abstract

Multi-view subspace clustering (MVSC) has acquired satisfactory clustering performance since it effectively integrates the information from multiple views. However, existing MVSC methods often suffer from high time costs and are difficult to be used in real-life large-scale data. Anchor-based MVSC methods have been presented to select crucial landmarks to reduce time-consuming effectively. In addition, the processes of anchor selection of existing methods are performed in the raw space, in which the high-dimensional data often involve lots of noise information and outliers that inevitably lead to the degradation of clustering performance. Moreover, these methods also ignore the balance structure of data, such that the selected anchors cannot fully characterize the intrinsic structure information of the original data. To tackle the aforementioned issues, we present a novel MVSC method named Fast Multi-view Subspace Clustering with Balance Anchors Guidance (FMVSC-BAG). Specifically, FMVSC-BAG integrates the learning processes of anchors, anchor graphs, and labels into a united framework in embedding space seamlessly. This way, they can reinforce each other to improve final clustering performance while eliminating noise and outliers hidden in the original data. Furthermore, FMVSC-BAG constrains the learned labels to preserve the balance structure by a novel balance strategy to promote further that the intrinsic balance structure information of original data can be reserved in the learned anchors and anchor graph. Finally, extensive experiments on eight real-life large-scale datasets prove its efficiency and superiority compared to some advanced clustering methods.