Skip to main navigation Skip to search Skip to main content

Bidirectional Adaptive Multilevel Locality Learning for Multi-View Graph Clustering

  • Ying Wang
  • , Jitao Lu
  • , Penglei Wang
  • , Haotian Li
  • , Danyang Wu
  • , Yamin Han
  • , Hongming Zhang
  • , Feiping Nie
  • Northwest Agriculture and Forestry University
  • University College Dublin
  • South China University of Technology
  • Shanghai Jiao Tong University

Research output: Contribution to journal › Article › peer-review

Abstract

Existing graph-based multi-view clustering methods usually learn clustering structures from graph similarity matrices or spectral embeddings separately, which ignores the interaction between the two procedures and weakens the exploitation of complementary graph information. Besides, most of them rely on global view fusion, while cluster-specific local preferences hidden in heterogeneous views are not fully explored. To alleviate these problems, a Bidirectional Adaptive Multilevel Locality Learning (BAML) model is proposed for multi-view graph clustering. Specifically, BAML first learns multilevel localities to provide adaptive guidance for bidirectional fusion, where dimension-level spectral weights identify informative spectral features, and two types of view-cluster weights characterize local preferences on the embedding and graph sides, respectively. With these learned localities, a parameter-free bidirectional fusion strategy is developed to integrate spectral representations and graph structures into a unified discrete clustering framework, enabling the two procedures to interact and complement each other for clustering. Moreover, an efficient alternating optimization algorithm is derived to solve the involved problem, and the corresponding convergence and complexity analyses are provided. Extensive experimental results on benchmark datasets demonstrate the effectiveness and superiority of BAML.

Keywords

  • Multi-view graph clustering
  • adaptive weight learning
  • bidirectional fusion
  • multilevel locality learning

Fingerprint

Dive into the research topics of 'Bidirectional Adaptive Multilevel Locality Learning for Multi-View Graph Clustering'. Together they form a unique fingerprint.

Cite this