Large Language Model (LLM)-enabled Graphs in Dynamic Networking

TL;DR

Dynamic networks demand adaptive, scalable optimization under time-varying conditions. The paper introduces a framework that integrates large language models with graphs, treating LLMs as predictors, encoders, and aligners within a layered architecture to enable graph-text and text-graph interchanges. It provides a taxonomy of LLM-enabled graphs, a practical five-layer framework, and a UAV-based case study demonstrating that LLM+GNN-based approaches can outperform baselines in trajectory planning and resource allocation. The proposed approach offers a path toward multimodal, real-time network optimization with broad applicability to UAVs, IoT, and other dynamic systems, and highlights future directions in adaptive optimization, security, and automation.

Abstract

Recent advances in generative artificial intelligence (AI), and particularly the integration of large language models (LLMs), have had considerable impact on multiple domains. Meanwhile, enhancing dynamic network performance is a crucial element in promoting technological advancement and meeting the growing demands of users in many applications areas involving networks. In this article, we explore an integration of LLMs and graphs in dynamic networks, focusing on potential applications and a practical study. Specifically, we first review essential technologies and applications of LLM-enabled graphs, followed by an exploration of their advantages in dynamic networking. Subsequently, we introduce and analyze LLM-enabled graphs and their applications in dynamic networks from the perspective of LLMs as different roles. On this basis, we propose a novel framework of LLM-enabled graphs for networking optimization, and then present a case study on UAV networking, concentrating on optimizing UAV trajectory and communication resource allocation to validate the effectiveness of the proposed framework. Finally, we outline several potential future extensions.

Figures (6)

• Figure 1: The processes and applications of LLM-enabled graphs: from graphs to text and text to graphs. LLMs play a crucial role in transforming graphical representations into textual information and converting text back into graphs. This dual capability facilitates deeper analysis and understanding of complex systems, such as power transmission and road networks.
• Figure 2: The summary of LLM-enabled graphs in different domains and dynamic networks. The applications of LLMs technology have led to important results in drug discovery of molecule design, and personalized marketing of E-commerce. Besides, LLMs have significant potential in dynamic networks and are expected to bring significant improvements to dynamic network application scenarios.
• Figure 3: The summary of LLM-enabled graphs and their applications in dynamic networks from the perspectives of using LLMs as predictors, encoders and aligners.LLMs as predictors can handle serialized graph structures or vectors generated by GNNs. As encoders, LLMs transform text information of nodes and edges into vector representations, enriching GNNs with textual data representations. Additionally, LLMs as aligners generate text embeddings that coordinate with the structural embeddings produced by GNNs, facilitating alignment between text and graphs.
• Figure 4: The structure of the proposed LLM-enabled graphs framework. The framework is based on a layered architecture consisting of an input layer, a graph-to-text layer, a decision layer, a text-to-graph layer, and an output layer. The input layer receives requests related to a dynamic network graph. The graph-to-text layer employs prompt engineering to extract features from requests, and then converts them to text via embeddings. The decision layer utilizes a pluggable LLM to generate responses. The text-to-graph layer extracts features from the text generated by the LLM to construct the graph, which is then processed by a GNN. The output layer analyzes the generated results and interacts with the user.
• Figure 5: The scenario module presents a system model of the considered scenario. The two-stage joint optimization strategy includes the first stage module and the second stage module. Specifically, an LLM is used to generate a UAV trajectory in the first stage, followed by employing a GNN in the second stage for communication resource allocation. The three modules interact with each other to obtain a UAV trajectory and communication resource allocation that meet the objectives and constraints.