Federated Node Classification over Graphs with Latent Link-type Heterogeneity.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Federated Learning

Background:

• Federated learning (FL) trains global models from decentralized data, but non-IID (non-independently and identically distributed) data poses challenges.
• Graph data often exhibits link-type heterogeneity, where links have varying semantics and homophily, differing across clients.

Purpose of the Study:

• To propose a novel graph FL framework that simultaneously discovers latent link-types and models link-specific message-passing.
• To address the challenge of link-type heterogeneity in graph federated learning.

Main Methods:

• Developed FedLit, a federated learning framework for graphs.
• Employed an EM-based clustering algorithm for dynamic latent link-type detection.
• Utilized multiple convolution channels to differentiate message-passing based on discovered link-types.

Main Results:

• Synthesized realistic graph datasets with latent heterogeneous link-types.
• Partitioned datasets to simulate varying levels of link-type heterogeneity.
• Demonstrated superior performance and rational behavior of the FedLit framework through comprehensive experiments.

Conclusions:

• FedLit effectively handles link-type heterogeneity in graph federated learning.
• The framework shows promise for improving FL performance on complex, real-world graph data.