Deep Fusion Clustering Network With Reliable Structure Preservation.

Area of Science:

• Machine Learning
• Data Mining
• Graph Neural Networks

Background:

• Deep clustering methods leverage data representation for sample partitioning.
• Combining auto-encoders (AE) with graph neural networks (GNNs) shows promise by incorporating structural information.
• Existing methods struggle with noisy graph connections, dynamic information fusion, and exploiting multi-view information for robust clustering.

Purpose of the Study:

• To propose a novel deep clustering method, DFCN-RSP, addressing limitations in existing approaches.
• To enhance the reliability of graph structures and improve representation learning.
• To achieve superior clustering performance through dynamic fusion and multi-view information exploitation.

Main Methods:

• Introduced a random walk mechanism to measure localized structure similarities, filtering noisy connections and preserving reliable ones.
• Developed a transformer-based graph auto-encoder (TGAE) with self-attention for refining fused topology.
• Implemented a dynamic cross-modality fusion strategy combining TGAE and AE representations, alongside triplet self-supervision.

Main Results:

• DFCN-RSP demonstrated superior performance compared to state-of-the-art deep clustering algorithms.
• The method effectively handles noisy graph structures and integrates attribute and structural information.
• Experimental results on five benchmark datasets validate the proposed approach's competitiveness.

Conclusions:

• DFCN-RSP offers a robust and effective deep clustering framework.
• The proposed techniques for structure preservation and information fusion lead to improved clustering accuracy.
• The method provides a competitive alternative for complex graph-based clustering tasks.