Graph Neural Network-Based Multicast Routing for On-Demand Streaming Services in 6G Networks

Basic Information

• Paper ID: 2510.11109
• Title: Graph Neural Network-Based Multicast Routing for On-Demand Streaming Services in 6G Networks
• Authors: Xiucheng Wang, Zien Wang, Nan Cheng, Wenchao Xu, Wei Quan, Xuemin (Sherman) Shen
• Categories: cs.NI (Computer Networks), cs.LG (Machine Learning)
• Publication Date: October 13, 2025
• Paper Link: https://arxiv.org/abs/2510.11109
• Code Link: https://github.com/UNIC-Lab/GNN-Routing

Abstract

With the growth of bandwidth-intensive applications in 6G wireless networks, such as real-time volumetric video streaming and multi-sensory extended reality, intelligent multicast routing solutions are required to deliver differentiated Quality of Service (QoS) at scale. Traditional shortest-path and multicast routing algorithms either incur excessive computational costs or employ rigid structures that often fail to support heterogeneous user requirements, resulting in poor resource utilization. While neural network-based methods offer improved inference speed, they typically lack topological generalization capability and scalability. To address these limitations, this paper proposes a Graph Neural Network (GNN)-based multicast routing framework that jointly minimizes total transmission cost while supporting user-specific video quality requirements.

Research Background and Motivation

Problem Definition

The core problem addressed in this research is the multicast routing optimization problem in 6G networks supporting heterogeneous QoS requirements. Specifically, it includes:

1. Heterogeneous User Requirements: Different users may require different video quality levels (from 360p to 8K) for the same content
2. Transmission Cost Minimization: Minimize total network transmission cost while satisfying all user requirements
3. Real-time Requirements: Provide low-latency routing decisions in dynamic network environments

Problem Significance

The development of 6G networks presents unprecedented challenges:

• Surge in Traffic Density: Holographic telepresence services require traffic density of 1-10 Tbps/km²
• Extremely High Data Rates: Real-time volumetric video applications may require peak data rates exceeding 100 Gbps per user
• Diverse QoS Requirements: XR applications involve synchronized audiovisual and tactile feedback, imposing strict requirements on reliability, latency, and throughput

Limitations of Existing Methods

1. Traditional Routing Algorithms:
   • Shortest-path algorithms such as Dijkstra and Bellman-Ford cannot exploit path reuse opportunities
   • Steiner tree-based multicast algorithms are NP-hard with excessive computational complexity
   • Assume homogeneous service requirements and cannot handle heterogeneous QoS demands
2. Neural Network Methods:
   • MLPs and CNNs require fixed input-output dimensions, lacking structural scalability
   • Poor generalization on unseen topologies
   • Cannot fully exploit the relational inductive bias of graph structures

Core Contributions

1. First Study: To the authors' knowledge, this is the first work investigating real-time video streaming multicast routing in 6G networks supporting differentiated user requirements
2. Problem Formulation: Formulates the multicast routing problem as a minimum cost flow optimization problem with inflow constraints, capturing both path reuse and user-specific QoS requirements
3. GNN Framework: Proposes a GNN routing framework based on graph attention mechanisms, achieving O(n) linear time complexity with generalization capability across arbitrary network topologies
4. Performance Verification: Extensive simulations validate the method's effectiveness, achieving near-optimal theoretical solutions while significantly reducing computational overhead

Methodology Details

Task Definition

Given a network graph G = (V, E), where V is the set of nodes and E is the set of edges. The network comprises:

• Source node set Vs (|Vs| = 1)
• Destination node set Vd (|Vd| = K)
• Relay node set Vr

Each edge (i,j) ∈ E has weight e(i,j) representing unit transmission cost. User demand vector x = x1, x2, ..., xK^T, where xk specifies the minimum required inflow for destination node k.

Optimization Objective:

min Σ(i,j)∈E e(i,j)f(i,j)

Constraints:

• Flow conservation constraints
• Demand satisfaction constraints
• Non-negativity constraints
• Topological feasibility constraints

Model Architecture

1. Theoretical Foundation

Theorem 1: Links carrying traffic form a tree structure rooted at the source node with all destination nodes as leaves.

Lemma 1: In the optimal solution, if a link is shared by multiple destination nodes, the traffic on that link equals the maximum demand among these destination nodes.

2. Policy Gradient Multicast Routing Method

Models route construction as a Markov Decision Process (MDP):

• State: st = (G, V(k)_inflow, Pt)
• Action: Select next-hop node vt
• Reward: rt = -x(k) * e(ut, vt)
• Objective: Maximize expected return ER

3. Graph Policy Network (GPN) Architecture

GAT Encoder:

eij = LeakyReLU(a^T[Wxi || Wxj])
αij = exp(eij) / Σk∈N(i) exp(eik)
hi = σ(Σj∈N(i) αijWxj)

LSTM Path Aggregator:

ht, ct = LSTM(xt; ht-1, ct-1)

Attention Decoder:

ptv = (ht-1)^T tanh(W2xv + W3ht-1)
πθ(st, at) = Softmax(ptv)

Technical Innovations

1. Structure-Aware Design: Leverages tree structure properties of optimal solutions to guide GNN design
2. Serialized Routing: Processes users in descending order of demand to achieve efficient path reuse
3. Attention Mechanism: GAT encoder learns importance weights between nodes
4. Memory Mechanism: LSTM captures sequential dependencies in routing decisions

Experimental Setup

Datasets

• Synthetic Network Topologies: Generated using NetworkX library
• Node Count: 30-50 nodes
• User Count: 1-15 users
• Connectivity: Fixed degree 3-6, average degree 3-7
• Demand Levels: High (1.0), Medium (0.5), Low (0.25)

Evaluation Metrics

• Transmission Cost: Total flow cost
• Execution Time: Routing computation time (logarithmic scale)
• Composite Score: 2×cost + log10(latency)

Comparison Methods

• Shortest Path Routing (Dijkstra)
• Genetic Algorithm (GA)
• Bee Colony Optimization (BCO)
• Dynamic Programming (DP): Theoretical optimal reference
• Graph Attention Network (GAT) Baseline

Implementation Details

• Hidden dimension H = 128, attention heads K = 4
• Adam optimizer, learning rate 5×10^-4
• Batch size 16, training for 20 epochs
• Gradient clipping threshold 1.0

Experimental Results

Main Results

1. Routing Cost Comparison

• Node Count Variation (30-50): GPN consistently outperforms GAT and Dijkstra, performs comparably to BCO, slightly higher than GA and DP
• Average Degree Variation (3-6): All algorithms show reduced cost with increased connectivity density, GPN maintains competitive advantage
• User Count Variation (1-15): GPN approaches theoretical optimum, significantly outperforms traditional methods

2. Time Complexity Analysis

Theorem 2: On sparse graphs, GA methods are at least Ω(GPU log|V|) times slower than GPN methods.

Experimental results show:

• GPN maintains sub-second execution time across all user counts
• Several orders of magnitude faster than GA, BCO, and DP
• Parameter count less than 3M, memory footprint under 50MB

3. Statistical Distribution Analysis

Violin plot analysis reveals:

• GPN exhibits compact low-cost distribution
• Small variance with good stability
• Distribution approaches theoretical optimum of DP

Ablation Studies

In 30-node 12-user scenario:

• Removing GAT: Significant transmission loss increase, demonstrating critical role of multi-head attention
• Removing LSTM: Slight performance degradation
• Removing Attention Pointer: Minimal impact

Dynamic User Addition

• GPN supports incremental re-routing, avoiding complete recomputation
• Maintains low transmission cost and rapid adaptation in dynamic scenarios

Related Work

Traditional Multicast Routing

• Wired Networks: DVMRP, MOSPF, PIM protocols
• Wireless Networks: MAODV, ODMRP and other mobility-adaptive protocols
• SDN Environment: Centralized control for dynamic optimization

Machine Learning Methods

• Deep Reinforcement Learning: Constructing multicast trees adaptive to dynamic topologies
• Metaheuristic Algorithms: Genetic algorithms, ant colony optimization for multi-objective optimization
• Graph Neural Networks: Emerging applications in network routing

Conclusions and Discussion

Main Conclusions

1. The proposed GNN framework effectively addresses heterogeneous QoS multicast routing in 6G networks
2. Achieves near-optimal performance while maintaining linear time complexity
3. Demonstrates good topological generalization capability and dynamic adaptability

Limitations

1. Static Weight Assumption: Assumes link weights remain stable during routing sessions
2. Snapshot Optimization: Optimizes based on latest measurements, requiring event-driven re-planning
3. Simulation Environment: Experiments primarily conducted on synthetic networks, lacking real network validation

Future Directions

1. Multi-source Multicast: Extension to multi-source scenarios
2. Multi-resolution Coordinated Scheduling: Coordinated scheduling of video streams
3. Practical Deployment: Deployment and validation in real 6G networks

In-Depth Evaluation

Strengths

1. Problem Importance: Addresses critical challenges in 6G networks with significant practical value
2. Theoretical Contributions: Provides theoretical analysis of optimal solution structure (Theorems 1 and Lemma 1)
3. Method Innovation: Cleverly combines GNN, reinforcement learning, and graph attention mechanisms
4. Comprehensive Experiments: Multi-dimensional comparative analysis covering cost, time, and scalability
5. Engineering Practicality: Low memory footprint suitable for edge deployment

Weaknesses

1. Insufficient Theoretical Analysis: Lacks convergence and approximation ratio guarantees
2. Experimental Limitations: Primarily validated on synthetic data, lacking real network scenarios
3. Incomplete Comparisons: Missing comparisons with latest deep learning routing methods
4. Simple Dynamic Handling: Relatively simple treatment of network dynamic changes

Impact

1. Academic Value: Provides new insights for GNN applications in network routing
2. Practical Value: Offers feasible solutions for 6G network multicast routing
3. Reproducibility: Open-source code provided for reproduction and extension

Applicable Scenarios

1. 6G Multimedia Services: Real-time video streaming, XR applications
2. Edge Computing: Intelligent routing in resource-constrained environments
3. Dynamic Networks: Network environments with frequent topology changes
4. Differentiated Services: Scenarios requiring support for heterogeneous QoS demands

References

The paper cites 43 references covering important works in graph neural networks, multicast routing, 6G networks, reinforcement learning, and other relevant domains, providing a solid theoretical foundation for this research.

Overall Assessment: This is a high-quality interdisciplinary research paper that successfully applies graph neural network technology to multicast routing in 6G networks. The paper demonstrates excellent performance in theoretical analysis, method design, and experimental validation, providing valuable solutions for addressing key challenges in future networks. Despite some limitations, its innovation and practicality make it an important contribution to the field.