Structure-aware federated hypergraph continual learning

Abstract

Federated hypergraph continual learning requires decentralized clients to learn a sequence of tasks over high-order relational data without directly sharing raw local data. This setting is challenging because hypergraph neural networks rely on hyperedges that encode relations among multiple nodes, while continual updates can gradually distort the representations learned from previous tasks. In this setting, existing continual preservation strategies mainly retain model parameters or output distributions, but they do not explicitly protect the node semantics and hyperedge structures that support old-task predictions. This paper proposes FedSHARP (Federated Structure-aware Hypergraph Anchored Representation Preservation), a structure-aware framework for federated hypergraph continual learning. FedSHARP treats forgetting as a multi-level phenomenon: output distributions may drift, node embedding centers may shift, and high-order hyperedge representations may lose their historical semantics. To address these issues, FedSHARP maintains a global structural memory bank that stores frozen task models, node prototypes, and hyperedge prototypes. During local training, multi-teacher knowledge distillation preserves old-task predictive behavior, node prototype alignment stabilizes class-level semantic representations, and K-Core-aware hyperedge anchoring gives stronger preservation to structurally central hyperedges. We evaluate FedSHARP on three hypergraph datasets, including CoauthorshipCora, CocitationCiteseer, and Yelp3K, under multiple federated task settings. FedSHARP achieves the best final average accuracy in all evaluated settings. In particular, on Yelp3K with hyperedge-based partitioning, FedSHARP improves final average accuracy from 0.3421 to 0.5491 compared with the strongest baseline. These results demonstrate that preserving both node-level semantics and high-order hyperedge structures is important for federated hypergraph continual learning.