An Adaptive Federated Relevance Framework for Spatial–Temporal Graph Learning

Spatial–temporal data contains rich information and has been widely studied recently due to the rapid development of relevant applications. For instance, medical institutions often use electrodes attached to different parts of a patient to analyse the electroencephalogram data with rich spatial and temporal features for health assessment and disease diagnosis. Existing research has mainly used deep learning techniques such as convolutional neural network (CNN) or recurrent neural network (RNN) to extract spatial–temporal features. Yet, it is challenging to incorporate both inter–dependencies spatial information and dynamic temporal changes simultaneously. In reality, for a model that leverages these spatial–temporal features to fulfil complex prediction tasks, it often requires considerable training data to obtain satisfactory model performance. To address the challenges at hand, we propose an adaptive federated relevance framework called FedRel for spatial–temporal graph learning. After transforming the raw spatial–temporal data into high–quality features, the core Dynamic Inter–Intra Graph (DIIG) module in the framework leverages these features to generate the spatial–temporal graphs capable of capturing both the hidden topological and long–term temporal correlation information. To further enhance the model's generalization ability and performance, while still prioritizing local data privacy, we have incorporated a relevance–driven federated learning module into our framework. This module leverages the diverse data distributions from different participants and applies attentive aggregations of respective models. We then conducted extensive experiments on two real–world spatial–temporal datasets. The results demonstrate the framework's efficacy in spatial–temporal interpretation, collaborative model training, and handling divergent data distributions across various comparison settings.