大语言模型文献挖掘驱动的网络指标体系与场景差异化分析

徐祺坤; 刘娅汐; 韩淑娴; 张慧峰; 皇甫伟

doi:10.11999/JEIT251120

大语言模型文献挖掘驱动的网络指标体系与场景差异化分析

doi: 10.11999/JEIT251120 cstr: 32379.14.JEIT251120

徐祺坤¹,
刘娅汐^1, ,,
韩淑娴^{1, 2},
张慧峰²,
皇甫伟¹

1.
北京科技大学计算机与通信工程学院北京 100083
2.
之江实验室杭州 311121

基金项目: 国家重点研发计划“网络模态共生演化模型与体系结构”(2023YFB2903901)

详细信息

作者简介:
徐祺坤：男，硕士生，研究方向为无线通信网络、多模态网络等

刘娅汐：女，副教授，研究方向为无线通信网络、6G通感一体化网络等

韩淑娴：女，硕士生，研究方向为人工智能、边缘计算等

张慧峰：女，高级工程师，研究方向为新型网络体系架构、智算中心网络等

皇甫伟：男，教授，研究方向为网络体系结构与理论模型、边缘计算等

通讯作者:
刘娅汐　yaxi.ustb@gmail.com

中图分类号: TN915.0
计量
- 文章访问数: 68
- HTML全文浏览量: 18
- PDF下载量: 6
- 被引次数: 0
出版历程
- 修回日期: 2026-06-29
- 录用日期: 2026-06-29
- 网络出版日期: 2026-07-08

Network Metric System and Scenario-Differentiated Analysis Driven by LLM Literature Mining

1.
School of Computer & Communication Engineering, University of Science and Technology Beijing, Beijing 100083, China
2.
Zhejiang Lab, Hangzhou 311121, China

Funds: 2023YFB2903901

摘要

摘要: 网络指标是网络设计、运维与优化的基础。本文提出一种基于大语言模型(Large Language Model, LLM)的文献挖掘方案，对《IEEE/ACM Transactions on Networking》2023-2024年刊载的583篇文献进行指标提取、术语规范化、场景标注与指标关联量化，并通过双LLM交叉验证和人工抽样复核检验挖掘结果。结果表明，网络指标呈现“少量共性指标+大量长尾专有指标”的结构，不同应用场景所关注的核心指标集及指标关联均具有场景依赖性。本文提出服务质量(Service, S)、资源复用成本 (Multiplexing, M)以及场景适配重构能力(Versatility, V)三类相互独立且可组合的分析维度，为网络架构的设计、部署、运维和优化提供指导。
- 网络指标 /
- 大语言模型 /
- 指标体系 /
- 多模态网络
Abstract: Objective Network metrics constitute the foundational parameters for network design, operation, and optimization. Existing research, however, predominantly focuses on single scenarios or representative parameters, lacking unified extraction rules and reproducible processes for large-scale, cross-scenario metric analysis. To address terminology ambiguity, scenario heterogeneity, and complex relationship quantification, a reproducible domain-specific literature mining chain driven by large language models (LLMs) is proposed. The objective is to automatically extract and standardize metrics, annotate application scenarios, quantify inter-metric dependencies, and establish an independent, combinable Service-Multiplexing-Versatility (SMV) theoretical framework. Rather than acting as a finalized suite of specific metric calculations, the SMV framework serves as a theoretical conceptual lens to guide multi-objective tradeoffs in architecture design and lifecycle management. Methods An automated literature mining framework is constructed utilizing a multi-agent LLM architecture (Fig. 1). A comprehensive dataset comprising 583 articles published in the IEEE/ACM Transactions on Networking (2023–2024) is analyzed. The methodology relies on three specialized agents: a terminology normalization agent mapping aliases and synonymous expressions into unified metric names; a scenario annotation agent assigning primary application tags based on high-density information sections; and a correlation mining agent identifying the semantic direction and intensity between metric pairs, quantifying these relationships into signed correlation coefficients ranging from -1 to +1. Mining reliability is verified through dual-LLM cross-validation and manual sampling review (Fig. 2). Results and Discussions The automated extraction identifies 3,978 independent network metrics, with 138 appearing in over 1% of the sampled literature (Fig. 3). The metric frequency distribution exhibits a pronounced heavy-tailed pattern, dominated by foundational parameters such as throughput (79.1%), end-to-end delay (74.6%), and packet error rate (Table 1). Focal metric sets demonstrate strong scenario dependence (Fig. 4). For instance, data center networks prioritize throughput and flow completion time, whereas Internet of Things (IoT) applications additionally emphasize energy consumption and network lifetime constraints (Table 2). Furthermore, inter-metric correlation matrices reveal highly differentiated coupling mechanisms across environments (Fig. 5, Fig. 6). In data center scenarios, throughput correlates negatively with flow completion time and queuing delay. Conversely, edge computing environments exhibit a negative correlation between end-to-end delay and resource utilization, reflecting an inherent tradeoff between real-time computing offloading latency and system resource efficiency. Conclusions The strong coupling between evaluation metrics and application scenarios necessitates a multidimensional perspective for future network architecture. Based on the extracted scenario-metric dependencies, the SMV analysis framework is conceptualized (Fig. 7). By integrating differentiated service quality requirements (Service), physical infrastructure overhead (Multiplexing), and adaptive reconfiguration agility for emerging services (Versatility), this paradigm enables adaptive resource orchestration in AI-native networks. Future research will focus on evolving the static literature mining pipeline into a dynamically updatable knowledge base and validating the SMV framework's engineering applicability in programmable and multimodal networks.
- Network metric /
- Large Language Model (LLM) /
- Metric system /
- Polymorphic network

HTML全文

图 1 LLM驱动的网络领域文献挖掘数据流图

下载: 全尺寸图片幻灯片

图 2 DeepSeek与GLM-4指标关联评分一致性散点图

下载: 全尺寸图片幻灯片

图 3 网络指标出现篇次排序及其分布

下载: 全尺寸图片幻灯片

图 4 不同网络场景出现的篇次及累计占比

下载: 全尺寸图片幻灯片

图 5 热点研究网络场景指标热力图

下载: 全尺寸图片幻灯片

图 6 非热点研究网络场景指标热力图

下载: 全尺寸图片幻灯片

图 7 网络指标体系的S-M-V三维框架示意图

下载: 全尺寸图片幻灯片

表 1 出现篇次前15位的核心网络指标

网络指标	出现篇次	篇次占比	定义
吞吐量	461	79.1%	单位时间内成功传输的数据总量
端到端时延	435	74.6%	数据从源端到目的端传输所需的总时间
误包率	347	59.5%	传输过程中错误数据包占总数据包的比例
丢包率	284	48.7%	传输过程中丢失数据包占总数据包的比例
误码率	222	38.1%	传输过程中错误比特数占总比特数的比例
带宽	215	36.9%	网络信道能够传输数据的最大速率
资源利用率	214	36.7%	网络资源(如带宽或处理能力)被使用的百分比
排队时延	144	24.7%	数据包在队列中等待处理的时间
抖动	121	20.8%	数据包到达时间间隔的变化量
往返时延	118	20.2%	数据从源端发送到目的端并返回所需的时间
信噪比	110	18.9%	信号功率与噪声功率的比值，用于衡量信号质量
传输时延	108	18.5%	数据从发送端到接收端传输所需的时间
跳数	98	16.8%	数据包从源到目的经过的网络节点数量
功耗	97	16.6%	网络设备在运行过程中消耗的功率
流完成时间	91	15.6%	一个数据流从开始传输到完成所需的总时间

下载: 导出CSV

表 2 分场景的网络核心指标特征

网络类型	重点关注指标(篇次占比)	场景特性描述
数据中心网络	吞吐量(86.9%)	大规模流量调度能力基准
	端到端时延(79.8%)	任务响应时延的关键约束
	流完成时间(63.6%)	拥塞控制的效能表征
边缘计算网络	端到端时延(79.2%)	包含传输和计算的总处理时延
	资源利用率(68.8%)	跨设备和服务器的负载分配优化
	通信开销(41.7%)	参数服务器与客户端之间通信所消耗的资源
物联网	吞吐量(76.7%)	单位时间内成功传输的传感器数据量
	端到端时延(66.7%)	传感器、执行器与边缘/云端反馈控制的实时性约束
	能耗(56.7%)	网络设备在处理和传输数据时的能量消耗
	网络生存周期(31.1%)	网络在能量耗尽前的持续运行时间
	采样率(27.8%)	传感器支持的每秒采集信号的次数
IP网络	丢包率(67.1%)	传输过程中丢失的数据包占总发送数据包的比例
	抖动(47.1%)	数据包到达时间间隔的变化量
移动自组织网络	吞吐量(76.2%)	动态路径的有效传输速率
	误包率(65.1%)	高扰动环境下的传输可靠性衡量
	信干噪比(44.4%)	高扰动环境下的无线信道稳定性表征
软件定义网络	吞吐量(90.6%)	VNF实例处理网络流量的能力
	实例扩展时延(81.1%)	根据负载动态调整规模的时间开销
区块链网络	吞吐量(85.7%)	网络单位时间处理的消息数量
	块传播时延(42.9%)	分布式共识效率瓶颈

下载: 导出CSV

参考文献(26)

[1]	WANG Chengxiang, YOU Xiaohu, GAO Xiqi, et al. On the road to 6G: Visions, requirements, key technologies, and testbeds[J]. IEEE Communications Surveys & Tutorials, 2023, 25(2): 905–974. doi: 10.1109/COMST.2023.3249835.
[2]	TANG Xiongyan, CAO Chang, WANG Youxiang, et al. Computing power network: The architecture of convergence of computing and networking towards 6G requirement[J]. China Communications, 2021, 18(2): 175–185. doi: 10.23919/JCC.2021.02.011.
[3]	任超, 丁思颖, 张晓奇, 等. 6G无线多模态通信技术[J]. 电子与信息学报, 2024, 46(5): 1658–1671. doi: 10.11999/JEIT231201. REN Chao, DING Siying, ZHANG Xiaoqi, et al. Wireless multimodal communications for 6G[J]. Journal of Electronics & Information Technology, 2024, 46(5): 1658–1671. doi: 10.11999/JEIT231201.
[4]	OBIODU E and SASTRY N. From ATM to MPLS and QCI: The evolution of differentiated QoS standards and implications for 5G network slicing[J]. IEEE Communications Standards Magazine, 2020, 4(2): 14–21. doi: 10.1109/MCOMSTD.001.1800041.
[5]	CAMPBELL A, COULSON G, and HUTCHISON D. A quality of service architecture[J]. ACM SIGCOMM Computer Communication Review, 1994, 24(2): 6–27. doi: 10.1145/185595.185648.
[6]	KARAKUS M and DURRESI A. Quality of service (QoS) in software defined networking (SDN): A survey[J]. Journal of Network and Computer Applications, 2017, 80: 200–218. doi: 10.1016/j.jnca.2016.12.019.
[7]	KAUSHIK N, KUMAR H, and RAJ V. A systematic review of QoS enhancement techniques in microservices architecture[J]. Computers and Electrical Engineering, 2025, 127: 110550. doi: 10.1016/j.compeleceng.2025.110550.
[8]	HANEMANN A, LIAKOPOULOS A, MOLINA M, et al. A study on network performance metrics and their composition[J]. Campus-Wide Information Systems, 2006, 23(4): 268–282. doi: 10.1108/10650740610704135.
[9]	AL-SHEHRI S, LOSKOT P, NUMANOGLU T, et al. Towards taxonomy of telecommunication network metrics[C]. Proceedings of 2017 European Modelling Symposium (EMS), Manchester, UK, 2017: 227–232. doi: 10.1109/EMS.2017.46.
[10]	PENG Cong, HE Yuanzhi, ZHAO Shanghong, et al. Integration of data center into the distributed satellite cluster networks: Challenges, techniques, and trends[J]. IEEE Network, 2023, 37(3): 52–58. doi: 10.1109/MNET.105.2100614.
[11]	XIAO Yang, SONG Yuqian, and LIU Jun. Collaborative multi-agent deep reinforcement learning for energy-efficient resource allocation in heterogeneous mobile edge computing networks[J]. IEEE Transactions on Wireless Communications, 2024, 23(6): 6653–6668. doi: 10.1109/TWC.2023.3335597.
[12]	张卫军, 蒋笑冰, 胡新, 等. 5G-R网络运维质量评价体系研究[J]. 铁道通信信号, 2025, 61(3): 1–9. doi: 10.13879/j.issn.1000-7458.2025-03.24031. ZHANG Weijun, JIANG Xiaobing, HU Xin, et al. Research on quality evaluation system of 5G-R network operation and maintenance[J]. Railway Signalling & Communication, 2025, 61(3): 1–9. doi: 10.13879/j.issn.1000-7458.2025-03.24031.
[13]	文浩, 林闯, 任丰原, 等. 无线传感器网络的QoS体系结构[J]. 计算机学报, 2009, 32(3): 432–440. doi: 10.3724/SP.J.1016.2009.00432. WEN Hao, LIN Chuang, REN Fengyuan, et al. QoS Architecture in Wireless Sensor Network[J]. Chinese Journal of Computers, 2009, 32(3): 432–440. doi: 10.3724/SP.J.1016.2009.00432.
[14]	ZHAI Yunkai, XU Xing, CHEN Baozhan, et al. 5G-network-enabled smart ambulance: Architecture, application, and evaluation[J]. IEEE Network, 2021, 35(1): 190–196. doi: 10.1109/MNET.011.2000014.
[15]	TAO Zhenyu, XU Wei, HUANG Yongming, et al. Wireless network digital twin for 6G: Generative AI as a key enabler[J]. IEEE Wireless Communications, 2024, 31(4): 24–31. doi: 10.1109/MWC.002.2300564.
[16]	GUO Daya, YANG Dejian, ZHANG Haowei, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning[J]. Nature, 2025, 645(8081): 633–638. doi: 10.1038/s41586-025-09422-z.
[17]	DAGDELEN J, DUNN A, LEE S, et al. Structured information extraction from scientific text with large language models[J]. Nature Communications, 2024, 15(1): 1418. doi: 10.1038/s41467-024-45563-x.
[18]	CHENG Xia, SHI Junyang, SHA Mo, et al. Revealing smart selective jamming attacks in WirelessHART Networks[J]. IEEE/ACM Transactions on Networking, 2023, 31(4): 1611–1625. doi: 10.1109/TNET.2022.3224358.
[19]	ZENG Yijing, CALVO-PALOMINO R, GIUSTINIANO D, et al. Adaptive uplink data compression in spectrum crowdsensing systems[J]. IEEE/ACM Transactions on Networking, 2023, 31(5): 2207–2221. doi: 10.1109/TNET.2023.3239378.
[20]	张慧峰, 胡宇翔, 朱俊, 等. 支持网络多模态共生与演化的体系结构及运行逻辑[J]. 电子与信息学报, 2026, 48(2): 607–617. doi: 10.11999/JEIT250949. ZHANG Huifeng, HU Yuxiang, ZHU Jun, et al. Architecture and operational dynamics for enabling symbiosis and evolution of network modalities[J]. Journal of Electronics & Information Technology, 2026, 48(2): 607–617. doi: 10.11999/JEIT250949.
[21]	WU Jiangxing, LI Junfei, SUN Penghao, et al. Theoretical framework for a polymorphic network environment[J]. Engineering, 2024, 39: 222–234. doi: 10.1016/j.eng.2024.01.018.
[22]	EBRAHIMI S, BOUALI F, and HAAS O C L. Resource management from single-domain 5G to end-to-end 6G network slicing: A survey[J]. IEEE Communications Surveys & Tutorials, 2024, 26(4): 2836–2866. doi: 10.1109/COMST.2024.3390613.
[23]	ZHANG Kaiyi, SAMAAN N, and KARMOUCH A. A machine learning-based toolbox for P4 programmable data-planes[J]. IEEE Transactions on Network and Service Management, 2024, 21(4): 4450–4465. doi: 10.1109/TNSM.2024.3402074.
[24]	WU Qiang, ZHAI X B, LIU Xi, et al. Performance tuning via lean measurements for acceleration of network functions virtualization[J]. IEEE/ACM Transactions on Networking, 2023, 31(1): 366–379. doi: 10.1109/TNET.2022.3193686.
[25]	MOHAMMADI A and NIKAEIN N. Athena: An intelligent multi-x cloud native network operator[J]. IEEE Journal on Selected Areas in Communications, 2024, 42(2): 460–472. doi: 10.1109/JSAC.2023.3336172.
[26]	朱俊, 徐琪, 张富军, 等. 面向云网融合环境的网络模态报文处理柔性流水线构建机制[J]. 电子与信息学报, 2026, 48(2): 522–533. doi: 10.11999/JEIT250806. ZHU Jun, XU Qi, ZHANG Fujun, et al. Flexible network modal packet processing pipeline construction mechanism for cloud-network convergence environment[J]. Journal of Electronics & Information Technology, 2026, 48(2): 522–533. doi: 10.11999/JEIT250806.