Generalized Category Discovery in Federated Graph Learning
Researchers have introduced a new framework called GCD-FGL to address limitations in Federated Graph Learning (FGL) regarding dynamic environments. Traditional FGL methods often rely on closed-world assumptions, failing to handle novel categories that emerge in decentralized graph data. This study targets Federated Graph Generalized Category Discovery (FGGCD), aiming to collaboratively identify new categories while preserving knowledge of known ones. The authors identify two primary challenges: the Neighborhood Absorption Effect, where structural fragmentation causes misclassification of novel nodes, and Global Semantic Inconsistency, where local biases are amplified across heterogeneous subgraphs. To mitigate these issues, GCD-FGL integrates a client-side Topology-Reliable Semantic Alignment process and a server-side Hierarchical Prototype Alignment strategy. Extensive experiments conducted on five real-world graph datasets demonstrate that the proposed framework consistently outperforms state-of-the-art baselines. The results show an average absolute gain of +4.86 in HRScore, indicating significant improvements in accurately discovering novel categories in distributed graph learning scenarios.
Wire timeline
Generalized Category Discovery in Federated Graph Learning
Researchers have introduced a new framework called GCD-FGL to address limitations in Federated Graph Learning (FGL) regarding dynamic environments. Traditional FGL methods often rely on closed-world assumptions, failing to handle novel categories that emerge in decentralized graph data. This study targets Federated Graph Generalized Category Discovery (FGGCD), aiming to collaboratively identify new categories while preserving knowledge of known ones. The authors identify two primary challenges: the Neighborhood Absorption Effect, where structural fragmentation causes misclassification of novel nodes, and Global Semantic Inconsistency, where local biases are amplified across heterogeneous subgraphs. To mitigate these issues, GCD-FGL integrates a client-side Topology-Reliable Semantic Alignment process and a server-side Hierarchical Prototype Alignment strategy. Extensive experiments conducted on five real-world graph datasets demonstrate that the proposed framework consistently outperforms state-of-the-art baselines. The results show an average absolute gain of +4.86 in HRScore, indicating significant improvements in accurately discovering novel categories in distributed graph learning scenarios.
cs.AI updates on arXiv.org