[PDF] GLASU: A Communication-Efficient Algorithm for Federated Learning with Vertically Distributed Graph Data-论文阅读讨论-ReadPaper

摘要

Vertical federated learning (VFL) is a distributed learning paradigm, wherecomputing clients collectively train a model based on the partial features ofthe same set of samples they possess. Current research on VFL focuses on thecase when samples are independent, but it rarely addresses an emerging scenariowhen samples are interrelated through a graph. For graph-structured data, graphneural networks (GNNs) are competitive machine learning models, but a naiveimplementation in the VFL setting causes a significant communication overhead.Moreover, the analysis of the training is faced with a challenge caused by thebiased stochastic gradients. In this paper, we propose a model splitting methodthat splits a backbone GNN across the clients and the server and acommunication-efficient algorithm, GLASU, to train such a model. GLASU adoptslazy aggregation and stale updates to skip aggregation when evaluating themodel and skip feature exchanges during training, greatly reducingcommunication. We offer a theoretical analysis and conduct extensive numericalexperiments on real-world datasets, showing that the proposed algorithmeffectively trains a GNN model, whose performance matches that of the backboneGNN when trained in a centralized manner.

展开全部

图表提取

论文十问

参考文献

被引用

提问与回答

暂无人提供速读十问回答