Federated graph-level clustering network with adaptive knowledge compensation.

Area of Science:

• Distributed graph computing
• Federated learning
• Machine learning

Background:

• Graph data is increasingly prevalent in real-world applications, necessitating advanced distributed computing solutions.
• Federated graph-level clustering frameworks show promise but struggle with personalized knowledge discrepancies among clients, hindering global model performance.
• Existing methods often fail to reconcile diverse client data, leading to suboptimal consensus and compromised clustering accuracy.

Purpose of the Study:

• To propose a novel federated graph-level framework designed to effectively mitigate personalized knowledge discrepancies in distributed graph clustering.
• To enhance the quality of client-side representations and ensure robust global model optimization through improved knowledge alignment.
• To achieve superior global consistency in federated graph clustering while preserving individual client performance.

Main Methods:

• Developed a Local Knowledge Enhancement (LKE) strategy on the client-side to extract and refine reliable, clustering-oriented representations using global prototype correction.
• Implemented a Global Prototype Alignment (GPA) mechanism on the server-side to construct affinity relationships and adaptively divide communities based on semantic similarity.
• Focused on optimizing knowledge alignment across clients under the principle of semantic similarity to enable effective global learning.

Main Results:

• The proposed framework demonstrated superior performance compared to existing state-of-the-art methods across multiple benchmark datasets.
• Achieved significant improvements in global consistency, indicating effective knowledge aggregation and alignment among clients.
• Successfully maintained high levels of personalized performance for individual clients, balancing global and local objectives.

Conclusions:

• The novel federated graph-level framework effectively addresses the challenge of personalized knowledge discrepancies in distributed graph clustering.
• The LKE and GPA strategies contribute to generating high-quality representations and optimizing global knowledge alignment.
• The framework offers a promising solution for enhancing federated graph clustering by improving both global consistency and personalized performance.