STFL: A Temporal-Spatial Federated Learning Framework for Graph Neural Networks

Therefore, it is of great significance to design such an end-to-end FL-GNN training framework. Graph Neural Network (GNN) has emerged recently owing to its ability to learn vector representations from complex In this paper, we propose a novel end-to-end spatialtemporal graph data. The GNN has now been widely used in applications federated learning framework for graph neural networks such as social network recommendation (Wu et al. (STFL), which can automatically transform time series 2021; He et al. 2019), traffic flow prediction (Wang et al. data into graph-structured data, and train GNNs collectively 2020; Cui et al. 2019), action recognition (Yan, Xiong, and to ensure good data privacy and model generalization. Lin 2018) and medical diagnosis (Rong et al. 2020; Sun We break our contributions into the following parts: et al. 2020). In addition to only using GNN models on learning 1. we first implement the graph generator to handle spatialtemporal the graph representations from different graph data, one data, including both feature extraction and node critical question is how to generalize the GNN models even correlation exploration; 2. we integrate the graph generator when the training data is insufficient, regardless of the graph into the proposed STFL and design an end-to-end federated or non-graph structures. This scenario applies to almost all learning framework for spatial-temporal GNNs on graphlevel cases where the data privacy is the major concern, and the classification tasks; 3. we perform extensive experiments model can only be trained to match the distribution of the on real-world sleeping dataset: ISRUC S3; 4. we publish local dataset.