On Inferring Communication Delays Using Semi-Supervised Learning

Abstract

This paper presents a graph-based neural network approach to estimating communication delays between node pairs from a limited number of known communication delays using semi-supervised learning. In recent years, graph-based machine learning models — such as GNN (graph neural network) and GCN (graph convolutional network), as well as their applications to network control and management problems — have been actively studied. These models can embed training data’s graph structure in their neural networks. The literature has proposed a communication delay estimation method between a reference node and other nodes using GCN that embeds a communication network’s topology. However, this method faces two limitations: it assumes a single reference node in the communication network, and it fails to explain how reference and measurement node placement affects estimation accuracy. In the current paper, expanding on this previous work, we propose a communication delay estimation method that utilizes communication delays between multiple reference nodes and a set of measurement nodes as training data. Also, through several experiments, we clarify the extent to which communication-delay estimation accuracy can be improved using communication delays from multiple reference nodes. Moreover, we show which locations in a communication network are suitable for placing reference and measurement nodes in order to estimate communication delays more accurately.