.png){kind=icon}
Return
Adaptive sparse graph for multi-view clustering
.png){kind=icon}
PRE
AI
DOI：10.1016/j.dsp.2026.105944.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Graph-based multi-view clustering (MVGC) has aroused interest as it can exploit consistent and complementary information from multiple perspectives. The quality of the constructed similarity graph largely determines the clustering performance of MVGC. Many existing methods directly apply the acquired similarity graph for spectral clustering, ignoring the massive inter-cluster similarities in the graph, influencing cluster partition. Constructing the k-nearest neighbors (KNN) sparse graph to remove inter-cluster similarities is a common improvement. However, kNN graph requires extensive tuning of the parameter k. To solve this, we propose a graph-based multi-view clustering method based on the adaptive sparse graph (MNV-MC). Specifically, an initial similarity graph is obtained by a low-rank tensor learning framework. Then, the heuristic method, Mutual Nearest Neighbor Value (MNV), is proposed to adaptively select the optimal k based on density changes to construct the high-quality sparse similarity graph. After processing by the fusion mechanism, the graph is input into spectral clustering to obtain clustering results. Experiments indicate that MNV-MC achieves outstanding performance, and the effectiveness of MNV for adaptively k-value selection of KNN graph is verified. Specifically, MNV-MC achieves average improvements of 7.79% in ACC and 5.16% in NMI over the second-best method across eight datasets, and gains of 7.29% and 5.79% on four additional large-scale datasets. Notably, as a parameter-free post-processing step, MNV can be easily integrated to other MVGCs. Experiments show that MVGC methods significantly improve their performance after applying MNV. The code is publicly available at https://github.com/ytccyw/MNVMC .
Journal
D
IF:
3
Papers:
779
Citations:
0
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSaveShareSaveShareSaveShareSave
View more