Federated Link Prediction on Dynamic Graphs
Abstract: Link prediction on dynamic, large-scale graphs has been widely used in real-world applications, such as forecasting customer visits to restaurants or predicting product purchases. However, graph data is often localized due to privacy and efficiency concerns. Training separate local models based on data in each region preserves privacy but often leads to less accurate models, especially in smaller regions with fewer users and products. Federated learning then collaboratively trains models on localized data to maintain model accuracy and data privacy. However, the vanilla FL approach requires training the entire historical graph of user interactions, introducing high computational costs during training. While training on the most recent data may help reduce overhead, it decreases the model accuracy and incurs data imbalance across clients. For instance, regions with more users will contribute more training data, potentially biasing the model toward those users. We introduce FedLink, a federated graph training framework for solving link prediction tasks on dynamic graphs. By continuously training on fixed-size buffers of client data, we can significantly reduce the computation overhead compared to training on the entire historical graph, while still training a global model across regions. Experiments demonstrate that FedLink matches the accuracy of training a centralized model while requiring 3.41$\times$ less memory and running 28.9% faster compared with full-batch federated graph training.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Jundong_Li2
Submission Number: 4330
Loading