Federated meta-learning for spatial-temporal prediction

Spatial-temporal prediction is a fundamental problem for constructing smart city, and existing approaches by deep learning models have achieved excellent success based on a large volume of datasets. However, data privacy of cities becomes the public concerns in recent years. Therefore, how to develop accurate spatial-temporal prediction while preserving privacy is a significant problem. To address this challenge, we propose a privacy-preserving spatial-temporal prediction technique via federated learning (FL). Due to inherent non-independent identically distributed (non-IID) characteristic of spatial-temporal data, the basic FL-based method cannot deal with this data heterogeneity well by sharing global model; furthermore, we propose the personalized federated learning methods based on meta-learning. We automatically construct the global spatial-temporal pattern graph under a data federation. This global pattern graph incorporates and memorizes the local learned patterns of all of the clients, and each client leverages those global patterns to customize its own model by evaluating the difference between global and local pattern graph. Then, each client could use this customized parameters as its model initialization parameters for spatial-temporal prediction tasks. We conduct extensive experiments on bike sharing datasets to demonstrate the superiority and effectiveness of our methods in privacy protection settings.