Tuesday, November 17, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Traditional Machine Learning Outperforms Large Language Models in 5G Traffic Classification

November 17, 2026
in Science News
Teresa Odom
By Teresa Odom Scienmag Editorial Profile - Machine Learning
Reading Time: 6 mins read
0
Traditional Machine Learning Outperforms Large Language Models in 5G Traffic Classification

Traditional Machine Learning Outperforms Large Language Models in 5G Traffic Classification

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Traditional structured machine learning models significantly outperform large language models when classifying operational fifth-generation mobile network traffic, according to an experimental investigation published in the Journal of Electrical Systems and Information Technology. Analyzing real-world cellular data across diverse application categories, researchers Mohamed E. M. Elhaj-Abdou and Waleed K. Badawi established that general-purpose language models struggle to process structured numerical telemetry, yielding high rates of prediction bias, output inconsistency, and reasoning errors. In direct comparative benchmarks, the evaluated open-weight models reached overall classification accuracies of merely 22.2 percent and 1.8 percent across ten behavioral traffic classes. These findings indicate that despite rapid advancements in natural language processing, modern autoregressive language architectures face substantial obstacles when deployed for automated network traffic characterization, falling far behind conventional machine learning algorithms that are inherently designed for tabular and statistical time-series data.

The research addresses a vital operational challenge stemming from the rapid global expansion of fifth-generation wireless infrastructure. Modern 5G cellular systems must simultaneously support diverse data-intensive and latency-sensitive applications, ranging from high-definition video streaming services like Netflix and Amazon Prime to real-time interactive communications such as Zoom and Microsoft Teams, alongside bursty traffic from cloud gaming platforms and metaverse environments. Each application family exhibits distinct quality-of-service mandates, packet arrival sequences, and throughput profiles, rendering legacy rule-based and port-based classification techniques obsolete due to widespread payload encryption, application-level obfuscation, and dynamic port allocation. Although automated traffic classification is essential for dynamic radio resource allocation and intelligent network management, earlier studies in cellular telemetry have overwhelmingly relied upon synthetic or simulated datasets. Such artificial traces consistently fail to capture genuine user dynamics, background interference, and the extreme statistical variability inherent to commercial carrier networks.

To overcome the documented shortcomings of simulated environments, the authors constructed a comprehensive real-world dataset gathered directly from operational mobile hardware over a six-month window from May to October 2022. The empirical measurement apparatus comprised a commercial Samsung Galaxy A90 5G smartphone outfitted with a Qualcomm Snapdragon X50 5G modem, communicating over the live commercial 5G production network of a major telecommunications operator in South Korea. Rather than deploying specialized packet sniffing hardware, the investigators captured device-level network packets using PCAPdroid, an open-source mobile inspection utility capable of logging packet traces without administrative root permissions. Data collection spanned more than 328 aggregate recording hours across stationary handsets executing specific applications, including Netflix, YouTube Live, Zoom, Roblox, and interactive cloud gaming. Background application activity was deliberately terminated during testing sequences to ensure each trace reflected the isolated behavioral properties of targeted services.

The acquired raw packet captures were subsequently converted into a structured time-series dataset through a standardized multi-stage preprocessing and sessionization pipeline. Captured packet headers were parsed to extract standard transport-layer descriptors, including source and destination IP addresses, port numbers, and protocol identifiers, allowing packets to be assembled into bidirectional communication flows. To capture temporal network dynamics, the researchers established an inactivity threshold of 60 seconds; any pause between consecutive packets exceeding this boundary initiated a new session. This sessionization procedure distilled the raw telemetry into 73,108 discrete traffic sessions. For every individual session, the engineering framework computed core key performance indicators: aggregate throughput, defined as total transmitted bytes divided by total session duration; packet density, representing the total observed packet count divided by session duration; and byte efficiency, measuring average payload utilization calculated as total transmitted bytes divided by total packet count.

Comprehensive statistical inspection of the 73,108 sessions demonstrated extreme numerical dispersion and heavy-tailed distributions typical of real-world wireless transmissions. The mean throughput across the dataset was measured at approximately 1.30 times ten to the third bytes per second, accompanied by a standard deviation of 1.76 times ten to the fourth bytes per second, while the interquartile range remained tightly compressed between 373 and 440 bytes per second. Session durations averaged 10,740 milliseconds with an enormous standard deviation exceeding 3.27 times ten to the fifth milliseconds, demonstrating the coexistence of brief micro-interactions and prolonged streaming connections. Similarly, while the median packet count stood at 18 packets per session, peak flows exceeded 1.5 times ten to the seventh packets. To stabilize numerical variance, researchers applied logarithmic transformations defined by log(1 + x) across duration, packet count, and throughput, alongside quantile-based rank normalization dividing each sample rank by total observations.

Using quantile thresholds derived across throughput, duration, and packet count, the authors mapped each session into one of ten semantic behavioral classes, including High Efficiency Flow, Persistent Congestion, Latency Dominated Flow, and Low Utilization Flow. Feature analysis across classes confirmed distinct operational signatures: High Efficiency Flow demonstrated the highest throughput and packet density, Persistent Congestion showed elevated session duration and packet volume reflecting transmission bottlenecks, while Latency Dominated Flow and Low Utilization Flow recorded subdued values across primary indicators. Scatter plots mapping log-transformed throughput against session duration revealed a general inverse relationship, where short sessions yielded high transmission rates while extended sessions trended toward lower throughput. However, the data also exhibited substantial clustering and overlapping boundaries between adjacent categories. These overlapping distributions demonstrated that even with engineered features, real-world cellular traffic exhibits intrinsic ambiguity that complicates algorithmic separation in multi-class environments.

To evaluate whether modern generative models could navigate this structured telemetry, the authors introduced an LLM-based classification framework comparing two state-of-the-art architectures: Qwen2.5-3B-Instruct and Gemma. The methodology transformed numerical session indicators into structured prompt representations combining a system prompt specifying task constraints and category definitions with a user prompt listing the specific performance metrics of a given session. Both models were executed within an identical GPU-accelerated software environment using half-precision sixteen-bit floating-point execution and deterministic generation parameters, setting sampling flags to false to guarantee reproducible evaluation. Inference was executed across all 73,108 sessions in the full experimental dataset. Following generation, an automated extraction pipeline parsed structured JSON outputs to isolate predicted class labels, employing fallback keyword matching and strict validation rules to handle unexpected syntax deviations or invalid model responses without corrupting the experimental evaluation.

Quantitative evaluation in the primary analysis layer exposed an immense performance disparity between the two language models, with both lagging substantially behind expectations for network automation. Gemma attained an overall classification accuracy of 0.222, a balanced accuracy of 0.294, and a macro-averaged F1-score of 0.168 across the ten traffic classes. In sharp contrast, Qwen demonstrated severe operational failure, recording an overall accuracy of just 0.018, a balanced accuracy of 0.117, and a macro F1-score of 0.009. Matthews correlation coefficient measurements remained near zero for both architectures, with Qwen producing a negative coefficient of -0.043, confirming an absence of reliable correlation between model predictions and ground-truth traffic behaviors. The study noted that pronounced dataset imbalance, characterized by the dominance of Severely Degraded Flow and Moderate Degradation Flow against sparsely populated minority classes like Unstable Throughput Flow, induced severe prediction bias that destabilized Qwen’s classification outputs.

The second evaluation layer examined computational efficiency and output stability, revealing critical operational hurdles for deploying generative architectures in real-time telecommunications environments. Gemma completed inference in an average of 3.05 seconds per traffic sample, whereas Qwen required 5.31 seconds per sample under equivalent computational configurations, highlighting a substantial processing overhead that limits practical real-time applicability. When evaluating prediction consistency across repeated or slightly varied inputs, Gemma recorded a consistency score of 0.83, reflecting relatively robust output repeatability. Conversely, Qwen produced a consistency score of only 0.07, demonstrating extreme vulnerability to minor input variations and erratic decision boundaries. This severe lack of stability directly explained Qwen’s poor classification accuracy, as inconsistent internal mappings prevented the model from maintaining generalized classification logic across structurally identical network sessions, illustrating how sensitivity to input formatting fundamentally undermines language model reliability on structured telemetry.

To interrogate reasoning reliability beyond aggregate statistical scores, the authors implemented a third evaluation layer using an LLM-as-a-judge protocol powered by DeepSeek across a representative sample of 300 randomly selected sessions. The DeepSeek judge evaluated model reasoning across four explicit dimensions: groundedness, contradiction, fabrication, and hallucination. Qwen demonstrated widespread reasoning failure, exhibiting hallucination and contradiction rates both exceeding 0.90, alongside a fabrication rate of approximately 0.66 and an average hallucination score of 0.89, proving that its textual rationales routinely contradicted provided metrics or invented non-existent network phenomena. Gemma performed substantially better, limiting its hallucination rate to approximately 0.39, its contradiction rate to 0.37, its fabrication rate to 0.06, and achieving an average hallucination score of 0.33. The DeepSeek judging mechanism introduced an evaluation latency of 2.19 seconds per sample for Qwen and 2.00 seconds for Gemma, demonstrating the feasibility of automated qualitative auditing.

The authors concluded that while large language models display emergent capabilities in conversational and unstructured reasoning, they remain fundamentally ill-suited for standalone structured 5G traffic classification in their default configurations. The research underscored that standard aggregate metrics like overall accuracy are insufficient for evaluating generative architectures in technical domains, necessitating multi-layer auditing that captures computational latency, prediction stability, and reasoning groundedness. To bridge the gap between structured network data and generative artificial intelligence, the investigators highlighted several essential avenues for future research, including the formulation of hybrid architectures that pair robust classical machine learning classifiers with language models acting as contextual interpreters. Additionally, developing specialized domain-adapted tokenization strategies, performing targeted parameter fine-tuning on cellular telemetry, and expanding benchmarking across wider commercial 5G and nascent 6G deployment environments will be crucial for creating reliable, carrier-grade network intelligence systems.

Subject of Research: Evaluation of machine learning and large language models for 5G network traffic classification

Article Title: Advancing LLM-based 5G traffic classification: a multi-layer evaluation of performance and reasoning reliability

Article References: Elhaj-Abdou, M. E. M., & Badawi, W. K. (2026). Advancing LLM-based 5G traffic classification: a multi-layer evaluation of performance and reasoning reliability. Journal of Electrical Systems and Information Technology, 13(1), Article 104. https://doi.org/10.1186/s43067-026-00402-5

Image Credits: AI Generated

DOI: 10.1186/s43067-026-00402-5

Keywords: 5G network traffic, traffic classification, large language models, machine learning, Qwen2.5, Gemma, DeepSeek, packet capture, PCAPdroid, quality of service, sessionization, inference latency, hallucination rate, cellular networks, network telemetry, packet density, byte efficiency, class imbalance

Cite Scienmag News

Teresa Odom. (November 17, 2026). Traditional Machine Learning Outperforms Large Language Models in 5G Traffic Classification. Scienmag. https://scienmag.com/traditional-machine-learning-outperforms-large-language-models-in-5g-traffic-classification/

Teresa Odom. "Traditional Machine Learning Outperforms Large Language Models in 5G Traffic Classification." Scienmag, 17 November 2026, https://scienmag.com/traditional-machine-learning-outperforms-large-language-models-in-5g-traffic-classification/. Accessed 17 November 2026.

Teresa Odom. "Traditional Machine Learning Outperforms Large Language Models in 5G Traffic Classification." Scienmag. November 17, 2026. https://scienmag.com/traditional-machine-learning-outperforms-large-language-models-in-5g-traffic-classification/

Tags: 5G network trafficbyte efficiencycellular networksclass imbalanceDeepSeekGEMMAhallucination rateinference latencylarge language modelsMachine learningnetwork telemetrypacket capturepacket densityPCAPdroidquality of serviceQwen2.5Sessionizationtraffic classification
Share26Tweet16
Previous Post

Feline Antibodies Enable Rapid Colorimetric Detection of Staphylococcus Aureus in Food

Next Post

Potassium Hydroxide and Calcium Lignosulfonate Triple Expansive Black Cotton Soil Strength

Related Posts

Religious-Aware Dataset and Neural Pipeline Improve Travel Document Head Covering Compliance Verification
Science News

Religious-Aware Dataset and Neural Pipeline Improve Travel Document Head Covering Compliance Verification

November 17, 2026
Refined TV Subtitles Substantially Improve Weakly Supervised Speech Recognition
Science News

Refined TV Subtitles Substantially Improve Weakly Supervised Speech Recognition

November 17, 2026
Polystyrene Microplastics Induce Anxiety via Amygdala Signaling Imbalance in Male Mice
Science News

Polystyrene Microplastics Induce Anxiety via Amygdala Signaling Imbalance in Male Mice

November 17, 2026
Tunable Flower-Shaped Terahertz Biosensor Achieves High Sensitivity for Cancer Cell Identification
Science News

Tunable Flower-Shaped Terahertz Biosensor Achieves High Sensitivity for Cancer Cell Identification

November 17, 2026
Digital Tax Adoption, IPSAS, and Tax Compliance Among Jordanian SMEs
Earth Science

Digital Tax Adoption, IPSAS, and Tax Compliance Among Jordanian SMEs

November 17, 2026
Chinese Visceral Adiposity Index Shows Nonlinear Link to Cardiovascular Disease Risk
Medicine

Chinese Visceral Adiposity Index Shows Nonlinear Link to Cardiovascular Disease Risk

November 17, 2026
Next Post
Potassium Hydroxide and Calcium Lignosulfonate Triple Expansive Black Cotton Soil Strength

Potassium Hydroxide and Calcium Lignosulfonate Triple Expansive Black Cotton Soil Strength

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Religious-Aware Dataset and Neural Pipeline Improve Travel Document Head Covering Compliance Verification
  • Refined TV Subtitles Substantially Improve Weakly Supervised Speech Recognition
  • Polystyrene Microplastics Induce Anxiety via Amygdala Signaling Imbalance in Male Mice
  • Tunable Flower-Shaped Terahertz Biosensor Achieves High Sensitivity for Cancer Cell Identification

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading