With advances in information acquisition technologies, multi-view data are increasing dramatically in a variety of real-world applications, whereas such data is usually corrupted by noises and outliers. Many existing multi-view graph clustering (MVGC) methods usually learn a consensus affinity graph using a late-fusion scheme in semantic space, which compound the challenge of leveraging the underlying relationships among corrupted multi-view data. In this paper, we propose a novel clustering method for handing corrupted multi-view data, hereafter referred to as Latent Low-Rank Proxy Graph Learning (LLPGL). Specifically, by projecting the multi-view data into a low-dimension proxy feature space, LLPGL can learn a low-dimension yet low-rank latent proxy from corrupted view data. Meanwhile, by employing the adaptive neighbor graph learning over the clean proxy, a high-quality affinity graph can be learned for clustering purpose. Then, an effective optimization algorithm is proposed to solve the model of LLPGL. Experimental results on five widely used real-world benchmarks validate the effectiveness of the proposed method.Consequently, the proposed method can be used to cluster the corrupted multi-view data for real-life applications.

随着信息获取技术的进步，多视图数据在各种实际应用中急剧增加，而这些数据通常会被噪声和异常值破坏。许多现有的多视图图聚类 (MVGC) 方法通常使用语义空间中的后期融合方案来学习共识亲和力图，这加剧了利用损坏的多视图数据之间的潜在关系的挑战。在本文中，我们提出了一种用于处理损坏的多视图数据的新聚类方法，以下称为潜在低秩代理图学习（LLPGL）。具体来说，通过将多视图数据投影到低维代理特征空间，LLPGL 可以从损坏的视图数据中学习低维低秩的潜在代理。同时，通过在干净代理上使用自适应邻居图学习，可以学习高质量的亲和图以进行聚类。然后，提出了一种有效的优化算法来求解LLPGL模型。五个广泛使用的现实世界基准的实验结果验证了所提出方法的有效性。因此，所提出的方法可用于对现实生活应用程序中损坏的多视图数据进行聚类。