Federated Learning Based On Graph Neural Network

With the vigorous development of Internet technology,the massive data accumulated by all walks of life has become one of the new driving forces for economic and social development.However,the accumulated data is often scattered,creating the problem of so-called data islands.In addition,reports of data abuse and leakage emerge in an endless stream,which has aroused people’s gradual attention to data privacy protection.Acts related to data privacy protection have also been introduced one after another.For example,the European Union has proposed the ”General Data Protection Regulation”,and my country has also proposed the ”Personal Information Protection Law of the People’s Republic of China”.Data supervision has become increasingly strict and comprehensive.Internet giants have also strengthened their control over data privacy and data security.Therefore,under the premise of ensuring data privacy,how to efficiently mine the value of data scattered on various data islands has become an important issue.Graphs,as a common data structure for modeling relations,have received extensive attention from academia and industry.The graph neural network has achieved good results in the utilization of graph data,such as financial anti-fraud,recommendation and other application scenarios.However,due to the existence of privacy protection and data island issues,there are many constraints on centralized data.However,only relying on the data owner with limited data may cause the model to overfit due to insufficient data,and it may be difficult to train a graph neural network model with sufficient accuracy.To sum up,it is of great significance to use federated learning to combine multiple data owners to jointly train the graph neural network under the premise of ensuring data privacy.In view of this problem,the existing work considers the training problem of graph neural network in the case of decentralization and the solution based on pseudo-nodes.However,there are two problems in these works: first,in the node-level federated graph learning problem,the potential connection problem between nodes across participants is not considered? second,the joint training of the graph and other heterogeneous data is not considered.situation.In order to deal with the problems of the above federated graph neural network,this paper proposes a corresponding improvement method,and the specific contributions are as follows:(1)A Federated Learning algorithm based on graph reconstruction is proposed to address the problem of potential missing connectivity across parties.:To complement potential connections across parties,this paper proposes a federated learning algorithm based on graph reconstruction.The method uses a node generator to predict that each node on each participant may have neighbor nodes across participants,and then uses a link prediction model to fill in the missing connections between newly generated nodes.This enables potential connections across participants to participate in the training update of the neural network.This paper verifies the effectiveness of the proposed federated learning algorithm based on graph reconstruction through experiments.(2)For the Federated Learning scenario composed of graphs and structured tables,this paper proposes a Federated Learning algorithm based on graph knowledge distillation.:In actual scenarios,some data owners do not have graph data to participate in model training,but only have structured tabular data.How to use graph data from other data owners to improve model performance through federated learning is a new problem.Therefore,this paper proposes a federated learning algorithm based on graph knowledge distillation,which can fuse graph data and structured tabular data to train the model in a federated learning manner.This paper verifies the feasibility and effectiveness of the method through experiments.In general,this paper studies the problem of potential connection loss of graph data across participants in the federated learning scenario,as well as the collaborative training problem of combining graph data and structured tabular data in the federated learning scenario,and proposes corresponding solutions.A series of experiments are designed to verify the proposed model,which shows the validity and rationality of the proposed model.