CGDD: Multiview Graph Clustering via Cross-Graph Diversity Detection.

Area of Science:

• Machine Learning
• Data Mining
• Graph Theory

Background:

• Multiview graph clustering is challenging due to noise and diversity within and across views.
• Existing methods often fail to adequately address diversity across multiple graphs, leading to corrupted similarity graphs.
• This limits the effectiveness of subsequent clustering tasks.

Purpose of the Study:

• To develop a robust multiview graph clustering method that explicitly addresses cross-graph diversity.
• To improve the recovery of clean similarity graphs by considering both intra-graph and inter-graph diversity.
• To achieve accurate clustering by fusing purified graphs into a consensus graph.

Main Methods:

• Proposed a novel approach focusing on sparse diversity across graphs, unlike previous methods concentrating on within-graph sparsity.
• Simultaneously detected multiview consistency and cross-graph diversity to purify individual view graphs.
• Fused purified graphs into a consensus graph with a predefined number of connected components (clusters).
• Employed an alternating iterative algorithm for adaptive similarity graph learning, diversity detection, graph fusion, and cluster assignment.

Main Results:

• The proposed model effectively recovers pure graphs for each view by leveraging cross-graph diversity.
• Fusion of these pure graphs resulted in a structured consensus graph accurately representing cluster structures.
• Experimental results on benchmark datasets demonstrated superior performance compared to state-of-the-art algorithms.

Conclusions:

• The method effectively handles noise and diversity in multiview graph clustering by focusing on cross-graph diversity.
• The approach leads to more accurate and robust clustering by generating a reliable consensus graph.
• This work offers a significant advancement in multiview graph clustering techniques.