.png){kind=icon}
.png){kind=icon}
Return
Parameter-free multi-view clustering via refined tensor learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.131497.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As multi-view data becomes more prevalent in real-world applications, multi-view clustering (MVC) has emerged as a powerful technique for unsupervised representation learning. To uncover the intrinsic structure, it is crucial to consider information from different spaces. Focusing solely on the sample space limits the method’s ability to effectively model multi-view data, as the informative patterns embedded in the feature space are often overlooked. Furthermore, to integrate high-order correlations, tensor-based MVC methods have been widely adopted to preserve the low rank structure of multi-view data. Traditional tensors can not achieve selective tensor rank minimization as they lack an explicit mechanism to model the retention of singular values based on their individual information contributions. Additionally, existing methods rely on hyper-parameters, undermining generalizability across different datasets. In response to these limitations, we propose a novel Parameter-free Multi-view Clustering via Refined Tensor Learning (PRTL), which is based on bidirectional regression matrices to perform data reconstruction and extract salient features. To further achieve an adaptive low-rank tensor structure, we propose a Quadratic Decay Tensor (QDT) regularization as a non-convex alternative to conventional rank minimization, which selectively retains salient information while filtering out noise dynamically, resulting in a more expressive joint representation. Meanwhile, we incorporate the hyper-Laplace graph to capture richer relationships than those modeled by conventional pairwise graphs. Notably, PRTL eliminates the need for hyper-parameters, making it more practical and robust. Experiments on diverse datasets demonstrate that PRTL consistently surpasses existing state-of-the-art clustering methods. Our code is available at https://github.com/jiaxinyang04/PRTL .
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W