Tensor-Representation-Based Multiview Attributed Graph Clustering With Smooth Structure.

Area of Science:

• Graph representation learning
• Machine learning
• Data mining

Background:

• Multiview attributed graph clustering shows promise with data augmentation.
• Multilayer graph autoencoders (GAEs) suffer from information aggregation deviation, especially with perturbed data.
• Existing methods are prone to bias from random view construction.

Purpose of the Study:

• To propose a tensor-representation-based framework (MV_AGC) for multiview attributed graph clustering that avoids bias.
• To enhance node representation learning and clustering stability.
• To improve overall clustering performance in attributed graphs.

Main Methods:

• Developed a tensor-product-based high-order graph attention network (GAT) with structural constraints for attribute fusion and semantic consistency.
• Integrated attribute augmentation and smooth constraints (SCs) into a high-order graph attention autoencoder.
• Introduced a clustering objective function-guided self-optimizing module to address performance degradation during clustering updates.

Main Results:

• MV_AGC effectively eliminates instability in reconstructed graph structures.
• The method learns more compact and robust node representations.
• Theoretical analysis demonstrates the superiority of the proposed tensor-product attention mechanism over classical GAT.
• Achieved state-of-the-art clustering performance on six benchmark datasets.

Conclusions:

• The proposed tensor-representation framework offers a stable and effective approach to multiview attributed graph clustering.
• MV_AGC demonstrates improved generality and expressiveness in node representation learning.
• The self-optimizing module further enhances final clustering accuracy, outperforming existing methods.