The core of their argument is a concept they call data endogeneity. In a classical machine learning pipeline, data is collected, cleaned, labeled, and then fed to a model; the model’s predictions do not change how the data was generated. In a learning-enabled 6G network, the opposite happens. The observations a learning system sees—channel measurements, traffic statistics, interference patterns, latency readings—are produced by the very communication, control, and architectural decisions the network itself makes. When a scheduler reallocates spectrum, the traffic statistics shift. When a beam-training algorithm changes its probing strategy, the channel measurements it later receives are different from what they would have been otherwise. The data is not a static resource sitting in a warehouse; it is a native byproduct of network dynamics, continuously regenerated by the system’s own behavior.

That feedback loop may sound like a philosophical nicety, but the survey shows it has hard technical consequences. Standard machine learning theory leans on assumptions such as independent and identically distributed samples, stationarity of the underlying process, and a clean separation between the training phase and the deployment phase. Wireless networks violate all three. The wireless channel is famously non-stationary: users move, obstacles appear, interference sources come and go, and the statistical properties of the signal environment drift on timescales from milliseconds to hours. Interference, in particular, exhibits temporal correlation under common fading models, meaning that successive observations are not independent draws from a fixed distribution but tightly coupled samples of an evolving process. A model trained on yesterday’s channel statistics may be stale by lunchtime.

Partial observability compounds the problem. No network element sees the whole system. A base station observes its own cells in detail and neighboring cells only dimly; a user device sees its local link but nothing of the broader topology. Learning algorithms deployed in such conditions must make decisions from incomplete, biased, and correlated views of the state, and the survey emphasizes that the placement of learning within the network architecture—what the authors call edge–cloud data locality—becomes a first-order design question rather than an implementation detail. Where data is aggregated, and where models are trained and updated, determines the statistical coherence of the learning problem and the timescales over which the network can actually adapt. A model refreshed every second at the network edge faces a fundamentally different learning problem from one retrained nightly in a central cloud, even if the underlying algorithm is identical.

Latency and resource constraints add a further layer of realism that much of the algorithmic literature ignores. 6G envisions services such as ultra-reliable low-latency communications, where decisions must be made within millisecond budgets and where failures carry safety implications for applications like industrial automation and vehicular control. Learning systems cannot simply pause the network while they retrain. Inference must run on hardware with finite compute and energy, often on edge devices, and the communication cost of coordinating distributed learning—federated updates, model synchronization, data shipment—competes directly with the network’s primary job of moving user traffic. The survey synthesizes work on edge intelligence, federated learning with non-IID data, and communication-efficient distributed training to map how these constraints reshape what is learnable in practice.

One of the most striking threads in the analysis draws on the emerging theory of performative prediction. In conventional supervised learning, the model is a passive observer of a fixed world. In performative settings, the model’s own predictions change the distribution it is predicting—a phenomenon first formalized in the machine learning literature and now, the authors argue, unavoidable in closed communication and control loops. A scheduling policy trained on observed traffic will alter that traffic once deployed; a reinforcement learning agent controlling radio resources changes the interference environment that its next observations will reflect. The survey reviews performative reinforcement learning in gradually shifting environments and state-dependent performative prediction, arguing that 6G networks are a natural habitat for these effects and that ignoring them leads to systematic, reproducible failures rather than random noise.

The closed-loop nature of learning-enabled networking also raises the stakes on stability and safety. When a learning controller sits inside a feedback loop with the physical channel and the network’s control plane, the question is no longer just whether the model’s accuracy is high but whether the coupled system remains stable. The authors survey work on learning-based model predictive control, predictive safety filters, control barrier functions, and safe reinforcement learning under partial observability—techniques developed largely in robotics and control theory that they argue must migrate into network design. Open RAN platforms, which expose programmable closed-loop control interfaces for machine-learning applications, are highlighted as both an opportunity and a testbed: they make it possible to study these dynamics experimentally, on real radio hardware, rather than only in simulation.

Perhaps the most uncomfortable part of the survey is its critique of evaluation practice. The authors examine benchmarking protocols commonly used in the field and identify systematic mismatches between those protocols and operational 6G reality. Static datasets stand in for live networks; independent test sets stand in for temporally correlated streams; offline metrics stand in for closed-loop performance. Time series forecasting research has long shown that naive performance estimation can badly mislead when data is non-stationary, and the survey argues that networking research inherits the same trap. The result is a recurring pattern of failure modes induced by abstractions: models that look excellent in the lab and degrade unpredictably in deployment, adaptation mechanisms that oscillate or diverge when coupled with the network, and benchmarks that reward exactly the assumptions real systems violate. The authors also invoke the broader machine learning literature on evaluation gaps and on data cascades in high-stakes AI, where the unglamorous work of data quality is skipped in favor of model work, with predictable consequences.

What the survey deliberately does not do is catalog algorithms. There are already numerous surveys enumerating machine learning techniques for resource allocation, network slicing, beam management, and physical-layer design. Trigka and Dritsas instead ask a prior question: under what system-level conditions does data-driven learning remain valid, stable, and interpretable inside an operational network? Their answer reframes Big Data and machine learning as endogenous system functions of 6G—components woven into the network’s own operation, subject to its physics, its latencies, and its feedback loops—rather than add-on intelligence layered on top. That reframing carries practical implications for how the field should evaluate claims, design experiments, and architect the AI-native networks that standards bodies and vendors are already building.

The timing is not incidental. Research programs on AI-native and task-oriented 6G architectures, foundation-model-based cloud–edge–end collaboration, and integrated sensing and communication are moving from vision papers to engineering specifications, and the decisions made in the next few years will harden into infrastructure that lasts decades. If the survey’s central claim is right—that the classical machine learning playbook quietly breaks when the learner and the learned system are the same machine—then the most important innovations in 6G intelligence may not be new architectures or exotic algorithms at all, but a more honest account of what happens when a network learns from itself. The authors, who contributed equally to the work, frame their contribution as a foundation for understanding, evaluating, and designing learning-enabled 6G systems beyond the algorithm-centric paradigms that have dominated the conversation so far. Whether the field listens may determine how much of the 6G hype survives contact with reality.

Subject of Research: System-level integration of Big Data and machine learning in 6G communication networks

Article Title: Big Data–driven machine learning for 6G communications: from algorithmic promises to system-level realities

Article References: Trigka, M., & Dritsas, E. (2026). Big Data–driven machine learning for 6G communications: from algorithmic promises to system-level realities. International Journal of Data Science and Analytics, 22(1), Article 292. https://doi.org/10.1007/s41060-026-01274-8

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01274-8

Keywords: 6G, machine learning, Big Data, wireless networks, edge intelligence, federated learning, non-stationarity, closed-loop control, performative prediction, O-RAN, network slicing, evaluation benchmarks