Information networks are essential for representing complex relational data, but privacy concerns often prevent the centralized collection of such data. While Federated Learning (FL) offers a privacy-preserving alternative, its application to graph-structured data faces two major hurdles: extreme data heterogeneity across different organizations and the dynamic nature of client contributions during the training process. Standard aggregation methods like FedAvg often fail to account for these variations, leading to suboptimal global models.
The FedSIN framework addresses these issues through a dual self-adaptive mechanism. Locally, it employs GAT to adaptively weigh the importance of neighboring nodes, allowing for more precise feature extraction tailored to each client’s specific network structure.On the server side, instead of using fixed weights, FedSIN integrates a Deep Deterministic Policy Gradient (DDPG) agent. This reinforcement learning model observes the state of local training rounds and dynamically allocates weights to each client,optimizing the global aggregation process in real-time.
Experimental results across various benchmarks demonstrate that FedSIN effectively handles Non-IID network data,providing a robust and scalable solution for multi-party information network analysis
