鲁棒核多视图子空间聚类

Robust Kernel Multi-View Subspace Clustering

摘要: 多视图子空间聚类旨在利用不同视图间的互补信息和一致性信息提升聚类性能。所提出的多视图子空间聚类方法是这一方向的重要工作，然而，该方法基于线性子空间假设，难以有效处理非线性多视图数据。为了解决该问题，本文通过核诱导映射将各视图数据投影至特征空间，使原本非线性的数据在特征空间中呈现线性子空间结构；同时，采用$l_{2,1}$ 范数替代Frobenius范数约束子空间表示误差，增强模型对噪声和离群点的鲁棒性。在多个真实数据集上的实验结果表明，所提方法在多个指标均优于对比方法，验证了核化策略的有效性。

Abstract: Multi-view subspace clustering aims to leverage the complementary and consensus information across different views to improve clustering performance. The proposed multi-view subspace clustering method is an important work in this direction. However, this method is based on the linear subspace assumption, making it difficult to effectively handle nonlinear multi-view data. To address this issue, this paper projects each view data into a feature space via kernel-induced mapping, enabling the originally non-linear data to exhibit a linear subspace structure in the feature space. Meanwhile, the $l_{2,1}$ -norm is adopted to constrain the subspace representation error instead of the Frobenius norm, thereby enhancing the robustness of the model to noise and outliers. Experimental results on multiple real-world datasets demonstrate that the proposed method outperforms the compared methods across multiple metrics, validating the effectiveness of the kernelization strategy.