Contrastive Multiview Attribute Graph Clustering With Adaptive Encoders.

Area of Science:

• Graph theory
• Machine learning
• Data mining

Background:

• Multiview attribute graph clustering leverages topological structures and node attributes.
• Existing methods struggle with inherent multiview graph relationships and sparse data.
• There's a need for methods that adapt to varying graph properties.

Purpose of the Study:

• To propose a novel contrastive multiview attribute graph clustering (CMAGC) method.
• To explicitly discover inherent relationships within and between multiview graphs.
• To effectively handle sparse graph structures in graph embedding learning.

Main Methods:

• Adaptive encoders are employed, selecting GCN encoder layers based on graph properties (e.g., high-order neighbors).
• Feature-level and cluster-level contrastive learning are applied to multiview soft assignments.
• Data augmentation uses the union of first-order neighbors as positive pairs to address sparsity.

Main Results:

• The proposed CMAGC method demonstrates superiority over state-of-the-art techniques.
• The approach effectively integrates multiview attribute graph information.
• Handles sparse graph structures and discovers inter/intra-view relationships.

Conclusions:

• CMAGC offers a significant advancement in multiview attribute graph clustering.
• The method provides a robust framework for handling diverse graph properties and sparsity.
• This work is the first to explicitly address inherent relationships from inter- and intra-view perspectives.