Multi-view graph-based clustering (MGC) aims to cluster multi-view data via a graph learning scheme, and has aroused widespread research interests in behavior detection, face recognition, and information retrieval in recent years. However, most of the existing MGC methods usually learn the affinity graph in the original space, such that they are inevitably hindered by the curse of dimensionality and corrupted features. Moreover, they usually learn the affinity between paired-samples by using Euclidean-distance metric; nevertheless, such a metric is sensitive to noise and outliers. In this paper, we propose a novel MGC method, namely latent embedding space learning (LESL), which aims to learn a latent embedding space and a robust affinity graph simultaneously. Specifically, a latent embedding representation is firstly learned, which can reduce the corruption and redundancy of the original views, and can effectively utilize the complementary information of multiple views. Afterwards, a robust estimator is used to automatically cut the connections among inter-cluster in the affinity graph. Finally, alternating direction minimization on the augmented Lagrangian multiplier (ALM-ADM) is adopted to optimize the unified objective function. Experimental results show that LESL outperforms state-of-the-art methods obviously.

多视图基于图的聚类（MGC）旨在通过图学习方案对多视图数据进行聚类，近年来在行为检测、人脸识别和信息检索方面引起了广泛的研究兴趣。然而，现有的大多数 MGC 方法通常在原始空间中学习亲和图，因此它们不可避免地受到维数灾难和损坏特征的阻碍。此外，他们通常使用欧几里德距离度量来学习配对样本之间的亲和力；然而，这样的度量对噪声和异常值很敏感。在本文中，我们提出了一种新的 MGC 方法，即潜在嵌入空间学习（LESL），旨在同时学习潜在嵌入空间和鲁棒的亲和图。具体来说，首先学习一个潜在的嵌入表示，可以减少原始视图的损坏和冗余，并且可以有效地利用多个视图的互补信息。之后，使用鲁棒估计器自动切断亲和图中集群间的连接。最后，采用增广拉格朗日乘子（ALM-ADM）上的交替方向最小化来优化统一目标函数。实验结果表明，LESL 明显优于最先进的方法。采用增广拉格朗日乘子（ALM-ADM）上的交替方向最小化来优化统一目标函数。实验结果表明，LESL 明显优于最先进的方法。采用增广拉格朗日乘子（ALM-ADM）上的交替方向最小化来优化统一目标函数。实验结果表明，LESL 明显优于最先进的方法。