Source-linked AI summary
Graph-based Deep Learning for Communication Networks: A Survey
Weiwei Jiang
TL;DR
Communication networks bring new operational problems as networks grow and new types emerge, while topology structure is not fully utilized by much deep learning. This survey comprehensively reviews graph-based deep learning applications in communication networks, reports effectiveness across a wide range of problems, and identifies future research directions, while noting training-data scarcity as a limitation.
Problem
Growing and changing communication networks create new problems, while much deep learning does not fully utilize network topology structure.
Method
The paper presents a comprehensive survey of graph-based deep learning applications in communication networks and maintains an open GitHub repository for updating relevant papers.
Results
Graph-based deep learning solutions are reported as effective for a wide range of communication-network problems and support generalization to unseen topologies.
Takeaways & Limitations
The survey organizes existing problems and solutions and identifies future research directions for graph-based deep learning in communication networks.
Takeaways & Limitations
Training data for graph neural networks remains a concern, with open training datasets still rare.
Abstract
from arXiv · showhide
Communication networks are important infrastructures in contemporary society. There are still many challenges that are not fully solved and new solutions are proposed continuously in this active research area. In recent years, to model the network topology, graph-based deep learning has achieved the state-of-the-art performance in a series of problems in communication networks. In this survey, we review the rapidly growing body of research using different graph-based deep learning models, e.g. graph convolutional and graph attention networks, in various problems from different types of communication networks, e.g. wireless networks, wired networks, and software defined networks. We also present a well-organized list of the problem and solution for each study and identify future research directions. To the best of our knowledge, this paper is the first survey that focuses on the application of graph-based deep learning methods in communication networks involving both wired and wireless scenarios. To track the follow-up research, a public GitHub repository is created, where the relevant papers will be updated continuously.
1. Introduction
Communication networks face traditional and emerging challenges, while conventional deep learning does not fully exploit network topology. This survey reviews graph-based deep learning applications across network scenarios, organizes prior studies, and identifies future directions.
- Motivation: Communication networks continue to develop through technologies such as SDN and SAGIN, creating traditional and emerging challenges including routing, resource allocation, and virtual network embedding.
- Motivation: Existing deep learning applications address networking problems, but most neural networks are designed for Euclidean data and therefore do not fully utilize network topology.
- Graph-based deep learning: Graph Neural Networks are proposed for non-Euclidean data and capture spatial information in network topology while generalizing to unseen topologies in dynamic networks.
- Survey scope: The survey covers graph-based deep learning studies across wireless, wired, and software-defined networks, with common problems such as network modeling, routing, and traffic prediction discussed across scenarios.
- Contributions: The authors summarize each study’s problem, graph-based solution, and GNN model, and provide a structured reference covering research from 2016 to 2021.
- Contributions: The paper proposes future research directions and maintains an open GitHub repository to update relevant studies for researchers entering the area.
2. Survey Methodology
The survey searches across graph-based deep-learning and communication-network keywords, major publishers, and multiple publication venues. It selects 81 papers and reports their annual and network-type distributions, with wireless networking receiving the most attention.
- Search strategy: The literature search combines graph-based deep-learning terms with communication-network and problem terms such as routing and traffic prediction.The graph-method terms include Graph Convolutional Network, Graph Attention Networks, GraphSAGE, Message Passing Neural Network, and Graph Isomorphism Network.
- Search strategy: The survey searches major publishers individually and uses Google Scholar to track citation relationships and reduce omissions from smaller publishers.Publishers listed include ACM, Elsevier, IEEE, Springer, and Wiley.
- Survey composition: Compared with 14 papers in 2019, the survey reports a 207% growth in papers in 2020, reaching 43 papers.The cited passage explicitly pairs the 207% growth with the increase from 14 papers in 2019 to 43 papers in 2020.
- Network-type distribution: Wireless-network studies receive more attention than wired and software-defined-network studies, and this trend may continue in 2021.The survey presents annual paper counts by network type in Figure 3.
3. Graph-based Deep Learning Introduction
This section introduces graph representations for wired and wireless communication networks and surveys graph-based deep learning models used with them. It covers graph embeddings, GCNs, message passing, and attention-based architectures, alongside benefits and limitations identified in the survey.
- Graphs in Communication Networks: Interference graphs model interactions between transceiver pairs, with undirected edges representing interference and edge features containing interference CSIs.The homogeneous ad-hoc example contains 3 nodes and 3 edges, while node features include direct CSI and environmental information.
- Graph Notation: The adjacency matrix incorporates topology into neural networks by encoding whether node pairs are connected, while symmetry distinguishes undirected from directed graphs.The section also introduces degree, Laplacian, and node-feature matrices for subsequent GNN definitions.
- Graph Neural Networks: GNNs extend neural networks to non-Euclidean data through message passing, updating node states by exchanging neighborhood information until reaching a stable state.The survey identifies GNN, GCN, and GAT as representative variants used in the reviewed studies.
- Graph Embedding: Graph embedding maps nodes, edges, and features into vector space while preserving graph structure and information as much as possible.Covered examples include Structure2vec, GraphSAGE, and GE; GraphSAGE learns an aggregation function that can generate embeddings for new nodes directly.
- Graph Convolutional Networks: GCNs generate node representations by aggregating a node’s features with those of its neighbors, using spectral-based or spatial-based formulations.Spectral GCNs operate through graph Fourier transforms, whereas spatial GCNs define convolution directly from graph topology; surveyed models include ChebNet and GCN.
- Advantages and Limitations: The survey notes that GNNs theoretically converge O(n log n) times faster and have O(n) times lower generalization error than multilayer perceptrons in communication networks with n nodes.It also reports that using more than two GCN layers can reduce performance, and that explaining GNN predictions remains an active area of progress.
4. Wireless Networks
The survey reviews graph-based deep learning applications across wireless network scenarios, including general wireless, cellular, and specialized networks. These studies address topology-aware optimization, prediction, resource allocation, and communication uncertainty.
- Wireless-network studies cover wireless local area, cellular, ad hoc, cognitive-radio, device-to-device, satellite, and vehicular networks.
- General Wireless Network: GNNs extract topology information for channel allocation, network-flow optimization, and other resource-management problems.GCN-extracted features can support deep reinforcement learning for channel allocation, while a topology-aware framework trades off computation time and inference performance.
- General Wireless Network: Wireless transmission uncertainty can deteriorate GNN robustness, motivating decentralized classifiers and retransmission mechanisms for power control and link scheduling.The mechanisms are considered for both uncoded and coded wireless communications.
- General Wireless Network: Power-control studies use REGNNs, UWMMSE, IGCNet, and MPGNNs to address allocation, interference, imperfect CSI, and beamforming.REGNNs outperform baselines with permutation invariance, while UWMMSE reduces computational complexity without harming allocation performance over a classic algorithm.
- Cellular Network: In cellular networks, GNN-based models improve multipath routing throughput and model spatial correlations for traffic prediction.The surveyed approaches include GAT, GCN, and GraphSAGE structures, alongside power-control, energy-efficiency, and idle-time-window prediction studies.
5. Wired Networks
The survey organizes wired-network applications around modeling, prediction, configuration, routing, management, and security. GNNs are used to represent network topology, estimate performance, support configuration decisions, and analyze traffic and alerts.
- Network Modeling: Wired-network modeling uses GNNs to represent existing and unseen topologies and estimate end-to-end metrics from topology, routing, and traffic inputs.Reported targets include delay, jitter, TCP throughput, and UDP latency, using supervised or semi-supervised approaches.
- Network Configuration: GNN-based methods support configuration feasibility, BGP synthesis, MPLS analysis, and network-property checking.DeepMPLS is designed to speed property analysis and suggest configuration changes when a property is not satisfied.
- Network Prediction: Traffic prediction studies combine graph models with temporal architectures and evaluate them on real backbone, ODTP, and science-network datasets.Examples include SGCRN, which combines GCN and GRU, and other graph-based models for real-world traffic prediction.
- Routing: Graph-based routing solutions address shortest-path, max-min routing, and load balancing while showing high accuracies and resilience to packet loss.The surveyed approaches include Graph-Query Neural Network and NGR frameworks.
- Network Security: Security applications use GNNs for botnet detection and intrusion-alert correlation, outperforming non-learning or traditional classification methods in the reported studies.Alert-GCN formulates intrusion-alert processing as node classification using an alert graph built from information from farther neighbors.
6. Software Defined Networks
The survey describes how graph-based deep learning supports SDN optimization by combining centralized performance measurements with topology-aware modeling. Applications include routing, bandwidth utilization, virtual-network embedding, resource management, and service-function chaining.
- SDN provides network-performance data for training GNNs, while GNNs use topology information to model and optimize networks.
- Routing Optimization: MPNN-based models estimate per-source/destination packet-delay distributions and loss with a worst Mean Relative Error (MRE) of 15.4%.These estimates are further used for routing optimization and network planning.
- Routing Optimization: A GN-based bandwidth-utilization solution achieves satisfactory accuracy with a prediction time 150 times faster than Genetic Algorithm (GA).
- Network Virtualization: Graph-based models obtain near-optimal virtual-network-embedding solutions in a short time and support VNF resource prediction.Reported applications include VNF forwarding-graph prediction, flow routing, flow scheduling, and topology management.
- Network Virtualization: GNN-based VNF methods report 90% average prediction accuracy, over 29% lower call-setup latency, and about 18% higher revenue-to-cost ratio in separate studies.The studies use real VoIP traffic traces or compare against simulated algorithms, respectively.
- Network Virtualization: Combining DRL with GCN improves acceptance ratio by up to 39.6% and average revenue by up to 70.6% over existing state-of-the-art solutions.
- Service Function Chaining: For SFC, GNNs outperform DNNs, while DRL-based approaches improve flexibility across new topologies without redesigning and retraining.The surveyed SFC literature is described as relying mostly on supervised learning, which may be unsuitable for changing resources, requests, and topologies.
7. Future Directions
The survey identifies future directions for combining GNNs with other AI methods, scaling models to larger networks, and addressing unresolved data and generalization challenges. It also characterizes graph-based deep learning for communication networks as an early research area with many open opportunities.
- Combining AI Techniques: Future work includes combining GNNs with GRU and DRL for traffic prediction, resource allocation, routing, and virtual-network embedding.
- Data and Model Development: GANs are suggested for settings with insufficient or expensive-to-collect training data, while combining GANs and GNNs remains unapplied in the survey’s communication-network scope.
- Data and Model Development: AutoML is identified as a way to optimize GNN parameters automatically.
- Larger Networks: GNN applications should extend to larger networks, because most surveyed topologies contain fewer than 100 nodes and larger models require substantial computation.
- Larger Networks: Graph partitioning and parallel computing are proposed for scaling, but the optimal divide-and-conquer strategy may vary across network scenarios.
- Larger Networks: The survey questions whether narrow performance margins justify the increased computational burden of graph-based models relative to traditional methods.
- Open Research Opportunities: The surveyed studies are described as only the beginning of graph-based deep learning research in communication networks, with a public GitHub repository for follow-up updates.
8. Conclusion
The survey organizes graph-based deep learning research in communication networks across wireless, wired, and software-defined scenarios. It lists each study’s problem and GNN-based solution, identifies future directions, and aims to serve as a reference for newcomers.
- The survey covers graph-based deep learning applications across wireless, wired, and software-defined network scenarios.
- For each study, it lists the problem addressed and the corresponding GNN-based solution.
- The paper identifies future research directions for follow-up research.
- The survey is intended to summarize recent progress and provide a reference manual for newcomers to the research area.