具有潜在平衡锚点指导的大规模多视图子空间聚类

Large-Scale Multi-View Subspace Clustering with Latent-Balance Anchors Guidance

摘要: 多视图子空间聚类(MVSC)是一种重要且广泛应用的技术，因为它能够有效地整合多视图信息并发现相关模式。然而，其高昂的计算成本限制了它在大规模数据集中的应用。为提高计算效率，基于锚点的MVSC方法应运而生，这些方法中选择锚点的策略包括直接对原始数据进行k-means聚类以获得锚点，或通过动态学习获得锚点。前者容易受到高维数据中的噪声和异常值的影响，从而导致聚类性能较差；后者主要依赖于正交性约束来提高锚点的多样性，但忽视了锚点的平衡结构和潜在语义关系，这可能削弱锚点的代表性和区分能力。为了解决这些问题，本文结合两种锚点选择策略的优点，提出了一种新的MVSC方法，称为带有潜在平衡锚点指导的大规模多视角子空间聚类。具体来说，本文将潜在完整空间学习、锚点搜索和锚点图构建整合到一个统一的框架中，实现相互加强和联合优化。与现有在原始空间中学习锚点和构建锚点图的方法不同，本文在潜在完整空间中进行锚点学习和图构建。此外，本文设计了一个新颖的正则化项，使锚点集向潜在质心对齐，从而在锚点学习过程中充分利用数据集的整体结构，使锚点集具有平衡结构并更具代表性。最后，在十个基准数据集上进行的大量实验证明了所提出的算法与现有最佳聚类方法相比所具有的有效性和优越性。

Abstract: Multi-view subspace clustering (MVSC) is an important and widely used technique because it can effectively integrate multi-view information and discover relevant patterns. However, its high computational cost limits its application to large-scale datasets. To improve computational efficiency, anchor-based MVSC methods have been proposed. These methods involve strategies for selecting anchors, including directly performing k-means clustering on the original data to obtain anchors, or dynamically learning anchors. The former is susceptible to noise and outliers in high-dimensional data, leading to poor clustering performance, while the latter mainly relies on orthogonality constraints to improve the diversity of anchors but ignores the balance structure and underlying semantic relationships, which may undermine the representativeness and discriminative power of the anchors. To address these issues, we combine the advantages of both anchor selection strategies and propose a new MVSC method called Large-Scale Multi-View Subspace Clustering with Latent-Balance Anchor Guidance. Specifically, we integrate latent space learning, anchor search, and anchor graph construction into a unified framework for mutual reinforcement and joint optimization. Unlike existing methods that learn anchors and construct anchor graphs in the original space, we perform anchor learning and graph construction in the latent space. Additionally, we design a novel regularization term that aligns the anchor set with the latent centroids, thereby fully utilizing the overall structure of the dataset during the anchor learning process, which results in an anchor set with a balanced structure and improved representativeness. Finally, extensive experiments on ten benchmark datasets demonstrate the effectiveness and superiority of the proposed algorithm compared to existing state-of-the-art clustering methods.