FedDGCL: Federated Graph Neural Network with Dual Graph Contrast Learning for Multivariable Time Series Forecasting

Ye Zhou, Yu Guo, Fangda Guo, Fangming Jing, Jiangrong Yang, Rongfang Bie

Abstract: Multivariate time series (MTS) forecasting plays an important role in various applications e.g., healthcare, economics, and traffic. Existing GNN-based methods generally utilize a predefined or learned graph to model the complex spatial dependencies between variables. However, these methods heavily rely on unrestricted access to centralized MTS data, which may bring data privacy concerns and a high cost of transferring data. Despite the emergence and success of federated learning (FL) as a decentralized training paradigm, making accurate MTS forecasting in FL is challenging due to the varied inter-variable relations among heterogeneous MTS data. In this paper, we propose a novel federated MTS forecasting framework named Federated Graph Neural Network with Dual Graph Contrast Learning (FedDGCL) to address the challenge. FedDGCL utilizes a dual graph learning module to decompose spatial dependencies as a shared graph for the universal part and a heterogeneous graph for the client-specific part. To ensure distinguishability and diversity of spatial dependencies, a novel graph contrast regularization method is designed to encourage the dual graphs to capture respective spatial information. Finally, FedDGCL introduces a dual temporal graph convolutional network to capture spatial-temporal dependencies. Experiments on five real-world datasets demonstrate that FedDGCL outperforms the state-of-the-art baseline methods while protecting data privacy.