Deep Learning on Graphs for Mobile Network Topology Generation

TL;DR

This work tackles pre-deployment generation of mobile network topology by predicting mobility relations between cells using inductive link prediction. It models the network as an undirected graph $G=(V,E)$ and compares a baseline MLP with a Graph Neural Network based on GraphSAGE to predict edge probabilities for a new node. Using PM/CM data from operational networks and ANR-derived edges for supervision, the study shows that incorporating graph structure improves prediction accuracy and precision, with the GNN typically outperforming the MLP. A geometry-based candidate filtering heuristic reduces training complexity and improves precision on imbalanced data, and the authors discuss limitations (e.g., bearing data absence) and future directions such as VGAE-based topology generation and additional attributes to enhance performance.

Abstract

Mobile networks consist of interconnected radio nodes strategically positioned across various geographical regions to provide connectivity services. The set of relations between these radio nodes, referred to as the \emph{mobile network topology}, is vital in the construction of the networking infrastructure. Typically, the connections between radio nodes and their associated cells are defined by software features that establish mobility relations (referred to as \emph{edges} in this paper) within the mobile network graph through heuristic methods. Although these approaches are efficient, they encounter significant limitations, particularly since edges can only be established prior to the installation of physical hardware. In this work, we use graph-based deep learning methods to determine mobility relations (edges), trained on radio node configuration data and reliable mobility relations set by Automatic Neighbor Relations (ANR) in stable networks. This paper focuses on measuring the accuracy and precision of different graph-based deep learning approaches applied to real-world mobile networks. We evaluated two deep learning models. Our comprehensive experiments on Telecom datasets obtained from operational Telecom Networks demonstrate the effectiveness of the graph neural network (GNN) model and multilayer perceptron. Our evaluation showed that considering graph structure improves results, which motivates the use of GNNs. Additionally, we investigated the use of heuristics to reduce the training time based on the distance between radio nodes to eliminate irrelevant cases. Our investigation showed that the use of these heuristics improved precision and accuracy considerably.

Figures (2)

• Figure 1: Block diagram of MLP.
• Figure 2: Block diagram of GNN. The upper part of the diagram produces the feature embeddings for the nodes, which are then used for the sampled possible neighbors as input to the lower diagram as in the MLP (figure \ref{['fig:MLP']}).