Through the telecom lens: Are all training samples important?

AI summary

Overview

• Research area: Machine learning for telecommunications — specifically data-efficient and energy-efficient training of telecom prediction models.
• Technical level: Intermediate. Readers should be comfortable with neural network training, loss gradients, and standard time-series metrics (MAE, RMSE); no advanced optimization theory is required.
• Scope: The paper studies whether individual training samples in telecom datasets contribute equally to model learning, and proposes a gradient-based sample importance framework that trains on a subset of data without losing accuracy.

What This Paper Is About

Most machine learning pipelines assume every training sample matters equally, an assumption inherited from curated image and NLP benchmarks. Telecom data is different: it is non-stationary, high-dimensional, costly to store, process, and label, and it reflects rare but operationally critical events. The authors ask whether training can be made faster and greener by identifying and training only on the samples that actually drive model improvement, and they test this on three real-world telecom datasets.

Key Contributions

1. Per-sample gradient dynamics analysis: An analysis of how individual sample gradients evolve across training epochs on a real-world open telecom dataset, showing that depending on dataset size, a significant portion of the training data contributes minimally to loss reduction or generalization.
2. A lightweight, model-agnostic importance mechanism: A method that estimates sample importance during training from gradient norm behavior and dynamically filters or re-weights data. Unlike influence functions or forgetting metrics, it requires neither retraining nor access to ground truth labels.
3. Empirical validation across three telecom datasets: Demonstrations on the Telecom Italia Big Data Challenge dataset, proprietary vendor data from 249 base station sites, and a synthetic 5G beam selection dataset that reduced-sample training matches or outperforms full-data baselines.
4. A sustainability angle: A CodeCarbon-based measurement showing that selective training lowers carbon emissions, tying AI training efficiency to telecom operators' energy and cost goals.

Main Findings

• Redundancy is real and grows with dataset size: The paper reports that the optimal proportion of data needed decreases as dataset size grows, indicating that larger datasets tend to exhibit redundancy.
• Telecom Italia results: Comparable performance to the full model is reached using 68% of important samples, which the paper reports as reducing training data by 28%.
• Vendor data results: 74% of samples suffice to match the baseline, reported as a 23% reduction.
• 5G beam selection results: The method outperforms the baseline on elevation angles using 75% of training samples, and matches the baseline on azimuth angles using 90% of the data.
• Prediction quality: In test prediction plots for the first two datasets, the full model and the best important-samples model are almost superimposed, and the framework also captures extreme drift in the Vendor data.
• Performance and time gains by dataset size (Table 1): At 5K samples, 80% of samples used, MAE improvement 2.71, training time improvement 3.07 seconds; at 50K samples, 70% used, MAE improvement 21.03, time improvement 15.55 seconds; at 300K samples, 65% used, MAE improvement 28.35, time improvement 21.03 seconds.
• Carbon emission reduction: The important-samples model consistently lowers emissions versus the full model — Telecom Italia from 2.1066 × 10⁻⁶ kg to 1.3031 × 10⁻⁶ kg (−38.14%), Vendor Data from 2.0734 × 10⁻⁶ kg to 1.2666 × 10⁻⁶ kg (−38.91%), and 5G Beam Selection from 1.9918 × 10⁻⁶ kg to 1.6927 × 10⁻⁶ kg (−15.02%), an average reduction of approximately 30.69%.
• Statistical robustness: Results are averaged over N_RUNS = 5 independent runs with error bars denoting mean ± 1σ, plus a bootstrap 95% confidence interval for the MAE improvement between the important-sample method and the full model.
• Small versus large datasets: Evaluations on LSTM-based models show small datasets yield limited gains, while larger datasets benefit substantially in training efficiency, computation time, and sustained or improved accuracy.

Methodology in Plain English

The researchers start from the idea that the update a model receives from a sample reveals how much that sample matters. For every training sample, at every epoch, they compute the norm of the gradient of the loss with respect to all trainable network parameters, and store these values in a matrix G of size E × N (epochs × samples). A sample's importance score is simply the average of its gradient norms across epochs.

They sort training samples by this score and select the top p% for retraining, sweeping p over {10, 20, ..., 90}. For each subset, the model is reinitialized and trained only on those samples, and MAE and training time are recorded and compared against a full-data baseline. The first two datasets use an LSTM benchmarked against prior art; the third uses the implementation and baseline from the original 5G beam selection paper.

The datasets span different telecom settings: the public Telecom Italia dataset (anonymized SMS, call, and Internet activity in Milan, here using the Internet_Activity feature as a univariate lag-1 forecasting task, normalized to [0, 1], split chronologically 80-20 with no shuffling to preserve temporal order); proprietary multivariate data collected every 15 minutes from 249 base station sites across LTE and 5G NR; and a synthetic 5G mmWave MIMO beam selection dataset built with the SUMO traffic simulator and the Wireless InSite ray-tracing tool, covering an urban canyon in Rosslyn, Virginia. Energy use was tracked with CodeCarbon.

Why This Matters

Impact on research: The paper challenges a baseline assumption in ML pipelines — that all training samples contribute equally — in a domain where data is expensive, private, and time-sensitive. It offers a gradient-based alternative to heavier techniques such as influence functions and core-set selection, and it is, to the authors' knowledge, the first study to apply an important-sample-based training strategy in the telecommunication domain. The authors frame it explicitly as a conceptual rather than algorithmic contribution, favoring empirical insight and practical utility over theory.

Real-world applications:

• Network traffic forecasting and KPI prediction models that must be retrained frequently as traffic patterns shift.
• Energy consumption modeling of base stations, where cell-level performance counters are aggregated to match hardware-level energy measurements.
• 5G mmWave beam selection, where azimuth and elevation angle prediction errors directly affect link quality.
• Anomaly detection and intelligent handover control in RAN operations, where retraining budgets are constrained.

Industry relevance: Telecom operators deploying AI across distributed infrastructure face continuous retraining under real-time or near-real-time constraints. The framework reduces redundant computation and data usage, which the paper links directly to operational cost and carbon footprint — relevant as networks scale toward dense 5G and 6G deployments and as initiatives like Green AI push for compute- and energy-efficient models.

Future Directions

1. Link gradient-norm statistics to generalization behavior, moving from empirical observation toward an explanation of why high-gradient-norm samples matter.
2. Incorporate dynamic or curriculum-style selection mechanisms, rather than a single static ranking computed from gradient norms.
3. Benchmark against more advanced core-set or influence-based approaches under comparable computational budgets to establish where the lightweight method wins and where it does not.
4. Validate across diverse model architectures and telecom tasks to assess how general the results are beyond LSTM-based forecasting and the beam selection setup.
5. Translate observed trade-offs into operational guidelines, such as how to choose the optimal data fraction under given latency or cost constraints — the paper states this remains an open area.

Target Audience

Researchers and engineers working on applied machine learning for telecommunications, particularly those responsible for model training pipelines, data curation, or sustainability targets in network operations. It is also useful for ML practitioners in any domain with large, noisy, time-ordered datasets who want a cheap alternative to influence functions or core-set selection, and for readers interested in the intersection of AI efficiency and energy consumption. Readers without background in neural network training and gradient computation will find the methodology section harder to follow.

Authors’ abstract

The rise of AI in telecommunications, from optimizing Radio Access Networks to managing user experience, has sharply increased data volumes and training demands. Telecom data is often noisy, high-dimensional, costly to store, process, and label. Despite Ai's critical role, standard workflows still assume all training samples contribute equally. On the other hand, next generation systems require AI models that are accurate, efficient, and sustainable.The paper questions the assumptions of equal importance by focusing on applying and analyzing the roles of individual samples in telecom training and assessing whether the proposed model optimizes computation and energy use. we perform sample-level gradient analysis across epochs to identify patterns of influence and redundancy in model learning. Based on this, we propose a sample importance framework thats electively prioritizes impactful data and reduces computation without compromising accuracy. Experiments on three real-world telecom datasets show that our method [reserves performance while reducing data needs and computational overhead while advancing the goals of sustainable AI in telecommunications.