Distributed Graph Neural Networks for Large-scale Social Network Data Analysis

Abstract

In light of the expansion in the scale and complexity of social network graph data, Distributed Graph Neural Networks (DGNNs) have been introduced as a new direction. To meet the demands of advanced systems, this paper proposes a new distributed graph neural network (DGNN) model that can effectively handle the diversity and dynamics of social graphs. This paper constructs an adaptive structure that can balance computational cost, communication expenses, and prediction accuracy in a distributed system. The system supports synchronous and asynchronous parallel execution, implementing its method through multi-level message passing and adaptive neighborhood sampling. The three public social network datasets used for the experiments are Friendster, Reddit, and Twitter, each containing millions of nodes and edges. According to the above experiments, the node classification accuracy of DGNN on all datasets is 0.80-0.88, with an F1 score exceeding 0.84. Superior to many other benchmark models. The scalability is significantly demonstrated in terms of reducing the training time from 500 seconds to 45 seconds per epoch and achieving over 85% parallel efficiency on a 16-node cluster. Robustness tests indicate that the accuracy of the graph significantly decreases in the presence of noise and missing edges. Research shows that the DGNN model excels in predicting resource efficiency and performance, demonstrating good generalization ability across various network environments. This paper will introduce a distributed graph learning system that can be used to analyze social networks.