⚡ electrical engineering

AI-Based KPI Prediction Methods in Future 6G Networks: A Survey

This survey presents the first comprehensive and systematic review of data-driven machine learning methods for predicting Key Performance Indicators (KPIs) in future 6G networks, offering a multi-dimensional taxonomy, analyzing state-of-the-art models, and outlining challenges and future research directions to enable proactive, AI-native network automation.

Original authors: Niloofar Mehrnia, Gourav Prateek Sharma, Samie Mostafavi, Andreas Johnsson, Sinem Coleri, Carlo Fischione, James Gross

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Niloofar Mehrnia, Gourav Prateek Sharma, Samie Mostafavi, Andreas Johnsson, Sinem Coleri, Carlo Fischione, James Gross

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the future of mobile networks (6G) not just as a faster internet, but as a self-driving car for data. Just as a self-driving car needs to predict where a pedestrian will step before they actually step, a 6G network needs to predict its own performance metrics before problems happen.

This paper is a massive "map" or "survey" of all the different ways researchers are teaching computers to make these predictions. Here is a breakdown of the paper's core ideas using everyday analogies.

1. The Problem: Why "Reactive" Isn't Enough

In the past, network managers were like firefighters. They would wait for a fire (a network crash, slow speed, or dropped call) to start, then rush to put it out. This is called "reactive" management.

But 6G networks are too complex and fast for firefighters. They need to be weather forecasters. They need to look at the clouds (data) and say, "It's going to rain in 5 minutes," so they can open the umbrellas (adjust resources) before the first drop falls. This is called predictive management.

2. The Four Things They Try to Predict (The KPIs)

The paper groups all the things a network needs to predict into four main categories, which the authors call KPIs (Key Performance Indicators):

Capacity (The Highway Width): How much traffic can the network handle? It's like predicting how many cars can fit on a highway during rush hour.
Latency (The Reaction Time): How fast does a message get from point A to point B? It's like predicting how long it takes for a letter to arrive. For things like self-driving cars, this needs to be instant.
Coverage (The Signal Blanket): Is the signal strong enough everywhere? It's like predicting where the Wi-Fi signal will be weak in your house so you can move the router.
Reliability (The Trust Factor): Will the message actually get through without errors? It's like predicting if a package will arrive intact or if it will get lost in the mail.

3. The "Toolbox" of AI Methods

The paper reviews hundreds of studies and organizes them by the "tools" researchers use to make these predictions. Think of these as different types of crystal balls:

Simple Math (Classical Models): Like using a basic calculator. Good for simple, steady patterns, but gets confused when things get chaotic.
The "Smart Learner" (Machine Learning): Like a student who studies past exams to guess future questions. It looks at history (past traffic, past weather) to guess the future.
The "Deep Thinker" (Deep Learning): Like a super-intelligent student who can see complex patterns humans miss. It can look at a map of the city and the time of day to predict traffic jams.
The "Group Think" (Ensemble Methods): Instead of asking one expert, you ask a committee of experts and take their average answer. This is often more accurate and stable.

4. The "Recipe Book" (The Taxonomy)

The authors created a new way to organize all these studies, which they call a taxonomy. Imagine a giant library where books aren't just sorted by "Fiction" or "Non-Fiction," but by:

What are they predicting? (Traffic? Signal?)
Where did they get the data? (Real-world measurements? Computer simulations?)
How far into the future are they looking? (Next second? Next hour?)
What tool did they use? (Simple math? Deep AI?)

This helps researchers see that some areas are well-studied (like predicting traffic), while others are empty (like predicting rare network failures).

5. The Real-World Hurdles (The "Gotchas")

The paper points out that while the math looks great on paper, putting it into real life is hard. Here are the main obstacles they found:

The "Black Box" Problem: Sometimes the AI gives a correct answer, but no one knows why. If a network manager can't understand the AI's reasoning, they are afraid to trust it. It's like a GPS that says "Turn left" but won't tell you why.
The "New City" Problem: An AI trained on traffic in New York might fail miserably in London. The models often don't work well when moved to a different location or a different type of phone.
The "Privacy" Problem: To predict well, the AI needs to know a lot about users (where they are, what they are doing). The paper warns that we need to be careful not to spy on people while trying to make the network faster.
The "Too Slow" Problem: If the AI takes too long to think, the prediction is useless. By the time it says "Traffic jam coming," the jam is already here. The prediction must happen in milliseconds.

6. The Future: What's Next?

The paper concludes that we are moving from "guessing" to "planning." The future isn't just about making the AI smarter; it's about making it safer, fairer, and more transparent.

Uncertainty: Instead of just saying "It will be slow," the AI should say, "It will likely be slow, but there is a 10% chance it will be fast." This helps the network prepare for the worst.
Self-Healing: The network should be able to notice when its own predictions are getting worse (because the world changed) and re-learn automatically.

Summary

This paper is a guidebook for the next generation of mobile networks. It tells us that to build a 6G network that runs itself, we need to stop waiting for problems and start predicting them. It reviews the best "crystal balls" (AI models) we have, organizes them into a clear system, and warns us about the traps (privacy, speed, and reliability) we need to avoid to make this vision a reality.

Technical Summary: AI-Based KPI Prediction Methods in Future 6G Networks: A Survey

1. Problem Definition and Motivation

The evolution from 5G to 5G-Advanced and the vision of 6G demand unprecedented network performance, requiring the support of emerging applications such as autonomous driving, industrial automation, and immersive communications. Traditional reactive network management is insufficient for these stringent requirements. Consequently, there is a critical shift toward AI-native architectures, where intelligence is embedded intrinsically across all network layers (RAN, core, and orchestration).

The core problem addressed is the need for predictive, data-driven approaches to forecast Key Performance Indicators (KPIs). Unlike reactive troubleshooting, AI-native networks require a "predictive sensing layer" to perceive state, anticipate future conditions, and proactively adapt behavior. This survey focuses on data-driven KPI prediction, defined as inferring an unknown KPI value ( $y_{t+\tau}$ ) from an available information set ( $I_t^k$ ) comprising historical measurements, counters, and contextual variables. The problem encompasses both estimation (inferring the current state, $\tau=0$ ) and forecasting (predicting future states, $\tau>0$ ).

The motivation for this survey stems from the fragmentation of existing literature. Prior works are often narrow-scope (focusing on a single KPI like throughput or coverage) or architecture-oriented (treating prediction as a sub-component of broader system discussions), lacking a unified taxonomy to compare methodologies across different KPI families, data sources, and prediction horizons.

2. Methodology and Taxonomy

The authors conducted a systematic review of 151 data-driven KPI prediction studies published between January 2000 and early January 2026. The methodology relies on a multi-dimensional taxonomy introduced to classify and compare these approaches along four orthogonal dimensions:

KPI Type: The survey categorizes targets into four families:
- Capacity: Throughput, goodput, and spectral efficiency at user, cell, or slice levels.
- Latency: End-to-end delay, round-trip time (RTT), and jitter, decomposed into RAN, transport, and core components.
- Coverage/Link Quality: Signal strength and quality metrics (Path Loss, RSRP, RSSI, SINR, CQI) and spatial coverage maps.
- Reliability: Error rates (BER, BLER, PER) and packet loss, often defined as successful delivery within a latency constraint.
Data Source: Classifies inputs as real measurement traces, simulated/emulated data, synthetic data, or hybrid combinations.
Prediction Attributes:
- Protocol Stack: The layer at which the KPI is predicted (PHY, MAC, RLC, PDCP, IP, Application).
- Horizon: Short-term (sub-minute) vs. Long-term/Static (hours to weeks).
- Output Formulation: Point estimates vs. probabilistic outputs (distributions, intervals).
Model Family: Ranges from classical statistical models (ARIMA, Linear Regression) and unsupervised learning (clustering, anomaly detection) to supervised learning (SVM, Random Forests), ensemble methods (XGBoost, LightGBM), and deep learning architectures (CNNs for spatial data, RNNs/LSTMs/Transformers for temporal sequences, GNNs for topology).

The survey analyzes the state of the art by mapping these studies onto the taxonomy, providing a quantitative landscape of the field and identifying methodological biases and gaps.

3. Key Contributions

The paper makes five primary contributions:

Unified Multi-Dimensional Taxonomy: It introduces a structured framework that enables consistent cross-study comparison, revealing that prior surveys often lack a holistic view of KPI dependencies and methodological trade-offs.
Quantitative Landscape Analysis: By annotating 151 studies, the authors provide the first empirical distribution of the field across KPI families, model types, data sources, and prediction horizons.
KPI-Centric Comparative Analysis: The survey offers a systematic review of prediction approaches for each of the four KPI families, detailing the evolution from classical statistical models to deep learning and reinforcement learning, and mapping methods to specific network domains (RAN, Core, Edge).
System-Level and Deployment Analysis: It examines the practical integration of predictors within 3GPP (MnS, NWDAF) and O-RAN (RIC, xApps) frameworks, addressing data collection, training pipelines, and operational constraints like telemetry granularity and edge resource limits.
Identification of Open Challenges and Roadmap: The authors outline critical limitations including reproducibility, evaluation metric inconsistency, offline-to-live integration gaps, inference-latency trade-offs, and the need for explainable and privacy-preserving predictors.

4. Results and Findings

The survey reveals several key trends and findings across the KPI families:

Capacity Prediction: While classical models (ARIMA, Linear Regression) serve as baselines, ensemble methods (Random Forests, XGBoost) are the industry standard for operational forecasting due to robustness and scalability. Deep learning (LSTM, GRU, Transformers) is increasingly used for capturing complex spatio-temporal dynamics, though generalization across spectrum bands and hardware configurations remains a challenge.
Latency Prediction: There is a shift toward tail-aware and distributional modeling (e.g., mixture density networks, quantile regression) to handle rare but critical latency spikes, particularly for URLLC. Sequence models (LSTM, Transformers) are dominant for capturing temporal dependencies, while Graph Neural Networks (GNNs) are emerging for network-scale inference involving routing and topology.
Coverage Prediction: The field has evolved from statistical path-loss models to multi-modal deep learning. CNNs and semantic segmentation networks (e.g., U-Net) are used to reconstruct Radio Environment Maps (REMs) directly from satellite imagery and 3D maps. Bayesian frameworks are increasingly adopted to decompose uncertainty (aleatoric vs. epistemic) for risk-aware planning.
Reliability Prediction: Predictors for error rates (BLER, PER) are moving from analytical formulas to learning-based approaches that combine physics priors with data-driven pattern recognition. Sequence models are effective for detecting bursty loss dynamics, and Reinforcement Learning (RL) is being integrated for proactive resource allocation to maintain reliability targets.

Systemic Observations:

Data Availability: Measurement-based studies dominate capacity and latency research, while reliability studies rely heavily on simulation due to the scarcity of labeled failure data.
Reproducibility: Only ~8% of studies provide public code, and ~20% provide data, hindering direct comparison and cumulative progress.
Evaluation Metrics: There is a lack of standardization; coverage/capacity studies often use point-error metrics (RMSE, MAE), while latency/reliability studies require tail-sensitive metrics (p95/p99, NLL) which are not always reported.

5. Significance and Claims

The paper positions itself as the first comprehensive and systematic review of data-driven KPI prediction methods specifically tailored for future 6G networks. Its significance lies in:

Bridging the Gap: It connects theoretical prediction algorithms with practical deployment realities in 3GPP and O-RAN ecosystems, highlighting the "offline-to-live" integration gap.
Roadmap for AI-Native Automation: It argues that accurate KPI prediction is the foundational enabler for zero-touch, intent-driven network operations. Without reliable prediction, autonomous control loops cannot function safely or efficiently.
Defining Future Directions: The survey explicitly calls for a shift from isolated accuracy improvements to robust, uncertainty-aware, and maintainable systems. It emphasizes the need for:
- Standardized benchmarks and open datasets to improve reproducibility.
- Probabilistic and explainable predictions to support risk-aware decision-making.
- Sustainable data lifecycle management to handle concept drift and model aging in dynamic 6G environments.

The authors conclude that while KPI prediction has matured significantly, the next frontier lies not in incremental accuracy gains, but in ensuring deployment robustness, generalization across domains, and safe integration into closed-loop control systems.