AI-Based Dynamic Spectrum Prediction and Allocation for IoT Wireless Networks Using Python

! ! ISSN!(Online):!2583-5696! Int.!Jr.!of!Hum!Comp.!&!Int.! ! ©!The!Author(s)!2025.!Open!Access!This!article!is!licensed!under!a!Creative!Commons!Attribution!4.0!International! License,!which!permits!use,!sharing,!adaptation,!distribution!and!reproduction!in!any!medium!or!format,!as!long!as! you!give!appropriate!credit!to!the!original!author(s)!and!the!source,!provide!a!link!to!the!Creative!Commons!license,! and!indicate!if!changes!were!made.!The!images!or!other!third!party!material!in!this!article!are!included!in!the!article’s! Creative!Commons!license,!unless!indicated!otherwise!in!a!credit!line!to!the!material.!If!material!is!not!included!in!the! article’s!Creative!Commons!license!and!your!intended!use!is!not!permitted!by!statutory!regulation!or!exceeds!the! permitted!use,!you!will!need!to!obtain!permission!directly!from!the!copyright!holder.!To!view!a!copy!of!this!license,! visit!http://creativecommons.org/licenses/by/4.0/! ! 637 RESEARCH!ARTICLE! OPEN!ACCESS! AI-Based Dynamic Spectrum Prediction and Allocation for IoT Wireless Networks Using Python G Nagendra Prasad . K Riyazuddin Department of Electronics and Communication Engineering, Annamacharya University, Andhra Pradesh, India. DOI:!10.5281/zenodo.17377265 Received:!17!September!2025!/!Revised:!11!October!2025!/!Accepted:!17!October!2025! ©Milestone!Research!Publications,!Part!of!CLOCKSS!archiving Abstract – The advent of the Internet of Things (IoT) has put pressure on scarce spectrum resources, especially in heterogeneous, interference-rich, and latency-critical environments. Static access regimes and rule-based ones cannot handle non-stationary interference and ultra-dense deployments. Leveraging advances in artificial intelligence (AI), this paper investigates and achieves spectrum intelligence: short-horizon spectrum prediction and dynamic, risk-aware allocation at radio timescales. We combine classical machine learning techniques, deep sequence and vision encoders (LSTM/GRU/TCN, spectrogram/REM models), transformers, reinforcement learning, and graph-based surrogates for channel–power assignment. Aside from modeling, reproducibility and deployability are also our focus with Pythonic pipelines: leakage-safe preprocessing, calibration (Brier/ECE), safety shields, and closed-loop evaluation with ns3/ns3-gym. Public datasets and simulators are inventoried with guidelines for splits and metrics (utilization, latency, violation rate, fairness, and energy). We cover engineering trade-offs for deployment on the edge (quantization, ONNX/TorchScript, federation with Flower) and outline open challenges in data sparsity, domain adaptation, explainability, security, and real-time feasibility. The result is an actionable roadmap to bring AI-enabled spectrum management from promising prototypes to robust, scalable IoT systems. Index Terms – AI-based spectrum management; IoT wireless networks; dynamic spectrum access; spectrum forecasting; reinforcement learning; deep learning; graph neural networks; Python implementation; cognitive radio; 5G/6G. I. INTRODUCTION The radio spectrum is the foundation of the Internet of Things (IoT), powering everything from low-power periodic sensing to latency-sensitive control-plane signalling. With billions of devices that are ! ! !!!!!!!!! ! 638 highly likely to run in heterogeneous environments, predicting when and where channels would be available has emerged as a pressing research issue. Early cognitive radio experiments demonstrated that simple machine learning (ML) baselines, trained from features such as duty cycle, frequency, and transmit power, could distinguish convincingly occupied versus free spectrum, establishing reproducible baselines for spectrum intelligence [1]. At the scale of smart cities, device diversity and bursty traffic, however, make demand highly volatile. Empirical evaluation on massive IoT traffic traces shows that good features and deep classifiers together can predict device behavior with extremely high accuracy [2]. This requires proactive spectrum prediction, where access policies are dynamically set downstream in anticipation of predictions of demand patterns in advance. But rule-based coexistence mechanisms and hard thresholds cannot realistically solve this issue. Experiments in illicit bands have shown that static approaches cannot handle non-stationary interference and stringent delay budgets, which present the necessity of acquiring learning-based control loops that can respond on radio timescales with coexistence guaranteed [3]. Recent advances in sequence learning enabled models that take advantage of temporal, spectral, and spatial dependencies within occupancy data. Composite recurrent forms, such as multi-dimensional LSTMs, provide principled approaches to encoding inter-step relationships for more precise multi-step forecasting at reasonable complexity [4]. Combined inter-channel relationships with temporal memory boost long-horizon prediction to allow anticipation-based reservation before bursts of contention [5]. Subsequent encoder-decoder transformer extensions with additional attention have since made prediction across multiple channels possible without the cost of prohibitive feature engineering, making it deployable in dynamic IoT environments [6]. Prediction alone is not sufficient; efficient allocation also has to be performed for handling the spectrum. Graph-neural surrogates have then emerged in the form of scalable near-optimal channel–power assignment approximators that balance accuracy and computation for dense IoT deployments [7]. Federated learning on Open RAN architectures presents an alternative path where distributed learning is performed without raw spectrum traces being transmitted out, hence eliminating privacy and nonstationarity concerns [8]. Reinforcement learning, particularly actor–critic methods, achieves explicit mapping from state space to spectrum allocation action, effectively closing the control–forecasting loop [9]. Other advancements, such as domain-knowledge-provided sequence models for geometric propagation smoothness representation [10] and distributed deep RL variants with good generalization under partial observability [11], bring us nearer robustness and practicability to real-time control. To this end, the focus of this paper is to investigate a general overview of AI-based approaches for spectrum prediction and allocation in IoT networks. Specifically, we will (i) compare the performance of multiple AI models in forecasting occupancy and demand in time, frequency, and space; (ii) elaborate how forecast data are integrated with adaptive allocation strategies; and (iii) introduce some recurring challenges such as the unavailability of datasets, explainability, and real-time applicability. Our three-fold contributions are: (i) to show a single, integrating pipeline that spans predictive models to allocation policies with IoT constraints; (ii) to analyse datasets, metrics, and implementation practices, pointing out strengths and weaknesses; and (iii) to introduce open research challenges, such as data scarcity, explainability, and real-time feasibility. The guidelines outlined here fill the gap between theoretical AI models and implementable spectrum management systems for massive IoT deployments. ! ! !!!!!!!!! ! 639 By building on prior cognitive radio intuition that merged prediction, sensing, and access [12], this paper builds an end-to-end spectrum intelligence overview to enhance large-scale IoT demands. II. SYSTEMATIC LITERATURE REVIEW Review scope and objectives The review charts the state of the art at the intersection of AI-driven spectrum prediction and dynamic spectrum allocation for IoT-oriented wireless networks, with a practical emphasis on Pythonic implementation and replicability. The review spans 2013–2025, from pioneering cognitive-radio work on prediction to current advancements for 5G and prospective 6G systems. The chapter tracks three questions: • which AI paradigms, traditional machine learning, deep learning, reinforcement learning, or hybrids, provide concrete gains for spectrum prediction and allocation, • what are data regimes, metrics, and evaluation practices • what obstacles still hinder deployment in heterogeneous, latency-sensitive IoT environments. Fig. 1: Year-wise count of the papers Identification Systematic searching in IEEE Xplore, ScienceDirect, ACM Digital Library, SpringerLink, arXiv, and Google Scholar was done using the following combinations of keywords: • "Spectrum Prediction" OR "Spectrum Occupancy Forecasting" OR "Radio Environment Map Prediction" • "Dynamic Spectrum Access" OR "Spectrum Sharing" OR "Channel Allocation" OR "Power Control" • "IoT" OR "Cognitive Radio Networks (CRN)" OR "5G" OR "6G" • "Machine Learning" OR "Deep Learning" OR "Reinforcement Learning" OR "Graph Neural Network" OR "Hybrid Models" • "Federated Learning" OR "Edge Intelligence" OR "Privacy-Preserving Spectrum Management" ! ! !!!!!!!!! ! 640 Screening and eligibility Screening occurred in two stages. Title–abstract screening retained papers that directly implied learningbased methods to spectrum prediction or dynamic allocation, or resources-optimization methods that were directly applicable to spectrum decisions (e.g., FK-LHSA, ADASA, RL-based DSS, hybrid RF–GRU– attention forecasting). Full-text analysis then confirmed the presence of an explicit algorithmic contribution and quantitative results (e.g., accuracy, throughput, latency, utilization, or energy). Inclusion criteria The final collection of primary studies was governed by the following criteria, modified to AI-based spectrum forecasting and dynamic assignment for IoT/CRN/5G/6G networks: • Time window: Appeared in peer-reviewed journals or flagship conferences between 2018–2025. • Problem focus: Focused spectrum forecasting (e.g., occupancy/demand/time-series forecasting) and/or dynamic selection/access/allocation of spectrum; tight cousin resource-optimization research was taken into account if the technique is tractably actionable for spectrum decisioning (e.g., FK-LHSA, ADASA, RL-based DSS). • AI methodology: Employed machine learning, deep learning, reinforcement learning, or hybrid pipelines; algorithms described with sufficient procedural detail to enable Python implementation (model class, features/inputs, training setup). • Quantitative evidence: Supplied task-relevant metrics, including at least one of: • Prediction: Accuracy/AUC/MAE/MSE, or error over multi-step horizons; • Allocation: Throughput, spectrum usage, collision/PU-violation rate, delay, fairness (e.g., Jain's index), or energy. • Data transparency: Utilized public measurements, well-delineated simulations (e.g., NS-3/custom with traffic/interference models), or open RF datasets (e.g., Kaggle/WiSig), with the dataset/split clearly documented. Fig. 2: Total count of the articles Data extraction and thematic synthesis The following fields were extracted from each included primary study to allow consistent analysis and reproducibility: • Dataset features: measurement vs. simulation, public/private availability, bands and bandwidth, sampling/observation window, dataset size, and train/validation/test protocol. ! ! !!!!!!!!! ! 641 • Model architecture: standard ML baselines (LR/SVM/RF/GBDT), temporal models (LSTM/GRU/TCN), Transformers for long-horizon prediction, CNN/vision backbones for spectrogram/REM inputs, RL (DQN/Actor–Critic/MARL), and graph-aware encoders (GNNs) when interference topology is represented. • Representation & features: hand-crafted radio features (RSSI, duty-cycle, cyclostationary/radiometrics), spectrogram/REM embeddings, attention-based representations, graph features, and environment/context features (load, QoS, mobility). • Training & evaluation setup: optimizer/schedule, horizon length for prediction, action space for allocation (channel/power/backoff), constraints (PU protection), and baselines/ablation definitions. • Performance metrics: prediction error (Accuracy/AUC/MAE/MSE), throughput, utilization, collision/violation ratio, latency, fairness, energy, and convergence (for RL). • Interpretability & diagnostics: feature importance (permutation/SHAP), saliency/attention maps on spectrograms/REMs, policy heat-maps and Q-value landscapes for RL, and failure-case analysis under distribution shift. • Explicit limitations: scale or realism of the dataset, compute/latency overhead, generalization between bands/topologies, safety guarantees, reproducibility gaps. The synthesis was organized around five core themes: • Shift towards sequence and Transformer-based predictors for multi-step, multi-channel prediction (LSTM/GRU/TCN - Transformer hybrids). • Simulation and data augmentation (NS-3/trace synthesis, domain adaptation from synthesized RF to real measurements) roles in data sparsity mitigation. • Modality and representation constraints, IQ streams, spectrogram/REM, CSI, coupled with preprocessing to make them model-ready. • Safety and explainability as first-class constraints, constrained or shielded RL and model attribution for auditability. • Generalizability and deployment readiness, such as cross-band/site transfer, federated/edge learning, scalability for multi-agents, and standard reporting. Reason and contribution of this review A fragmented literature has grown in which prediction and allocation are typically solved as separate problems, evaluated on non-comparable data with heterogeneous measures and minimal reporting of failure modes or constraints. This survey adopts a pipeline-based view: prediction elements that forecast occupancy/demand are conditionally linked to allocation elements tha decide channel/power, so that progress in one stage is evaluated in terms of total utility (utilization, latency, fairness, and PU safety). The threefold contribution. First, a methods taxonomy is created that covers classical optimization, deep temporal/vision models, reinforcement learning, and graph-aware encoders. Within this taxonomy, the review specifies when each family is useful (e.g., Transformers for long-horizon predictions; MARL for decentralized distribution; GNNs when interference topology prevails) and what they cost (data gluttony, compute/latency, reproducibility sensitivity). Second, the review formalizes report standards in support of strict replication and comparison in Python: fixed seeds and splits; disclosure of observation window and action space; baselines mandated (e.g., optimization-only or heuristic); and a minimum set of metrics ! ! !!!!!!!!! ! 642 (prediction error + {throughput, utilization, collision/PU-violation, latency, fairness, energy}). The standards align with operational constraints within IoT settings and seek to reduce historical imprecision that has stunted cross-paper synthesis. Third, the review values credibility. For prediction, this entails stress-testing under distribution shift (band/site/time) and publishing attribution analysis. For allocation, particularly RL, it encourages constraint handling (hard shielding or constrained objectives), safety audits against incumbent tampering, and policy interpretability through state-action summaries. The survey also catalogs data streams, from public RF corpora (e.g., Kaggle/WiSig) to NS-3 simulation roll-outs, so researchers can pretrain at scale and subsequently validate on realistic small-scale traces. Comprehensive Review The rapid progress in wireless communication networks, such as cognitive radio networks (CRNs), 5G, and next generation 6G systems, has caused crucial challenges in spectrum utilization, resource allocation, network traffic forecasting, and optimal service quality, requiring intelligent, adaptive, and predictive techniques. Spectrum prediction as a central module of cognitive radio systems has been studied extensively to improve spectrum sensing, decision-making, sharing, and mobility. Xiaoshuang et al. [13] provided a detailed survey of prediction methods for the spectrum in CRNs, emphasizing the need for accurate prediction to reduce processing latency and improve spectrum efficiency, and identifying open problems such as dynamic channel availability, real-time adaptability, and prediction accuracy in heterogeneous environments. Building on these concepts. Bharathi et al. [14] developed the AI-Assisted Adaptive Searching Algorithm (AIAS), which integrates supervised learning for predicting spectrum availability with reinforcement learning for realtime adaptive allocation to secondary users. Their method, tested on the Kaggle Spectrum Dataset, yielded 91.3% prediction accuracy, 92.5% average spectrum utilization efficiency, and interference-free allocations of up to 95.1%, cutting latency by about 50%, illustrating the real-world feasibility of AI in adaptive spectrum management in highly utilized environments. In the high-reliability and ultra-low latency of 5G networks that support the boom of IoT devices as well as ultra-dense deployments, Ramesh et al. [15] introduced the FK-LHSA scheme that combines Fractional Knapsack-based multi-band spectrum selection and Lagrange Hyperplane-based spectrum access allocation, optimizing the throughput, reducing spectrum access delay, and increasing the accuracy of allocation in IoT sensor networks. Adding to this, Banoth Ravi et al. [16] used machine learning algorithms to achieve intelligent, predictive, and adaptive spectrum allocation in 5G and beyond networks with real-time self-optimization and energy-efficient network resource management through accurate forecasting of spectrum demand. In cognitive communication paradigms for heterogeneous wireless networks, the application of AI has been shown to further enhance spectrum utilization. Kai Lin et al. [17] developed the AI-based Data Analyticsbased Spectrum Allocation (ADASA) algorithm, integrating deep learning for feature extraction, dimensionality reduction, and user correlation analysis, allowing adaptive parameter tuning according to the network scenario and achieving significant gains in spectrum efficiency. Similarly, Ramy et al. [18] proposed reinforcement learning-based Dynamic Spectrum Sharing (DSS) for next-generation networks, wherein the AI-bespoke framework selects resources dynamically based on existing network conditions and user activity, leading to enhanced LOS rates, throughput, spectrum efficiency, and latency minimization, while indicating continued challenges of data accessibility, algorithmic computational expense, security, and standardization. aside from spectrum management, ! ! !!!!!!!!! ! 643 Rawan et al. [19] explored the paradigm of personalized wireless networks with the introduction of an AIdriven, big data-powered, surrogate-assisted multi-objective optimization system that micro-manages scarce network resources at the individual user level with competing objectives of resource optimization and user satisfaction. The solution speaks to the feasibility of AI to provide real-time tailored service quality, opening smart network structures that move beyond stiff, preconceived QoS models. Network traffic prediction and QoS enhancement have also become critical in low-latency, high-speed environments. Ibrahim [20] addressed this by developing a stacking ensemble approach for network intrusion detection, backed by LSTM and LSTM Encoder-Decoder models for predicting 5G network QoS metrics like throughput, latency, jitter, and packet loss. Tested over a large-scale 50-day field dataset, the models achieved high attack detection rates (90.4% and 98.7%) with F1-scores of 90.0% and 98.5%, and low prediction errors for QoS metrics (14.57% and 13.75%), demonstrating AI’s effectiveness for simultaneous security and service optimization in live 5G environments. Scaling this to 6G networks, Mohammed Anis et al. [21] proposed a hybrid AI model of Random Forest (RF), combined with Gated Recurrent Units (GRU) and attention mechanisms, to accurately predict 6G network traffic in various channel conditions and user scenarios. The model achieved excellent performance metrics, including RMSE of 0.0049, MAE of 0.0034, MAPE of 0.46%, and R² of 0.9970, far surpassing baseline GRU and LSTM models, and enabling proactive resource planning, secure robustness, and real-time network optimization. Providing an overview, Karthick [22] surveyed AI methods applied in wireless communication such as machine learning, deep learning, and reinforcement learning to manage spectrum, allocate resources, detect signals, and predictive maintenance and identified open challenges like cross-layer integration, standardization gaps, computational overhead, and data privacy concerns, thereby giving a vision for the future of intelligent, self-organizing, and autonomous wireless networks. Limitations of Existing Studies Identified Despite the comprehensive development reported in employing AI for predicting the spectrum, allocating it, and optimizing wireless resources, several limitations have been identified in the existing studies, which restrict real-world implementation and scalability of these solutions. 1. Absence of Real-World Validation: Much of the work (e.g., Xiaoshuang et al. [13], Kai Lin et al. [17], Banoth Ravi et al. [16], Bharathi et al. [14]) is simulationor controlled-dataset-based (e.g., the Kaggle Spectrum dataset). Simulations provide information on algorithmic performance but do not capture the full richness of dynamic, heterogeneous, interference-rich IoT and 5G/6G deployments. This poses a problem with the generalizability of the models described to real deployments. 2. High Computational Complexity and Latency: AI-based models such as ADASA [17], reinforcement learning-based DSS [18], and RF–GRU–Attention hybrid models [8] provide high accuracy but at the cost of computational load. These approaches will likely not scale for ultradense IoT environments where real-time processing with ultra-low latency is an absolute requirement. Similarly, ensemble and hybrid techniques require intense training, which may not be feasible on resource-constrained IoT devices. 3. Scalability Issues in Dense IoT Networks: FK-LHSA [15] and AIAS [14] schemes are effective in small-to-medium scale applications but are bound to fail under extreme IoT device connectivity. ! ! !!!!!!!!! ! 644 Endless monitoring and re-allocation need create scalability bottlenecks that, accordingly, may cause delay, spectrum fragmentation, and increased interference under ultra-dense deployments. 4. Large and High-Quality Dataset Dependence: Various techniques (e.g., Rawan et al. [19], Ibrahim [20], Mohammed Anis et al. [21]) rely on big data and long-term datasets for training. In practice, it is expensive, time-consuming, and often impossible to acquire such datasets in dynamic spectrum environments. Furthermore, data sparsity, imbalance, and privacy concerns further limit the capability of data-driven AI models. 5. Limited Cross-Layer and Cross-Network Integration: Latest research aims at spectrum prediction or allocation on a single network layer only, without end-to-end integration between layers (physical, MAC, network) or among heterogeneous wireless systems (e.g., CRNs, 5G, 6G). According to Karthick [22], this does not allow for end-to-end optimization and hence creates gaps between spectrum management and overall network QoS, security, and energy efficiency. 6. Security, Privacy, and Standardization Challenges: Dynamic spectrum sharing frameworks (e.g., Ramy et al. [18]) also identify concerns of security threats, standardization deficiencies, and trust in AI-driven spectrum allocation. Since sensitive spectrum utilization information is processed by AI systems, concerns about adversarial attacks, data breaches, and the absence of standardized protocols for secure adoption persist. 7. Energy and Resource Constraints in IoT Devices: Whereas some articles (e.g., Banoth Ravi et al. [16]) focus on energy efficiency improvements, most of the existing frameworks assume high computational power and unlimited resources, which is unrealistic for low-power IoT devices. This incompatibility prevents the achievement of AI-based spectrum allocation in resourceconstrained and battery-limited IoT settings. 8. Restricted Investigation of Advanced Deep Learning Models: Despite conventional ML and hybrid methods dominating existing research, advanced deep learning architectures (e.g., Transformers, Graph Neural Networks) remain underexploited for spectrum prediction and assignment. Bharathi et al. [14] also note that the integration of deep learning can possibly improve robustness and accuracy but is more a future research direction rather than an adopted method Table 1: Overview of the existing literature Reference Model / Algorithm Strength Limitation [13] Survey of Spectrum Prediction Techniques Comprehensive overview of spectrum sensing, decision, sharing, and mobility; identifies key challenges and research directions No experimental validation; purely survey-based; lacks performance metrics [15] FK-LHSA (Fractional Knapsack + Lagrange Hyperplane) Optimizes multi-band spectrum selection and allocation; reduces access delay; improves throughput and accuracy Requires continuous monitoring; may not scale for extremely dense networks [17] ADASA (AI-driven Data Analytics-based Spectrum Allocation) Uses deep learning for feature extraction and data correlation; adaptive allocation based on network conditions; improves spectrum utilization High computational complexity; real-world deployment not demonstrated [18] AI-based Dynamic Spectrum Sharing (Reinforcement Learning) Real-time adaptation to network conditions improves LOS, throughput, spectrum efficiency, and reduces interference High computational cost; standardization and security issues remain [19] AI-enabled Big Datadriven Multi-Objective Optimization User-level personalization balances resource efficiency and user satisfaction; real-time adaptability Relies on large datasets; potential scalability issues for very large networks ! ! !!!!!!!!! ! 645 [16] ML-based Predictive Spectrum Allocation Real-time network self-optimization; energy-efficient spectrum management; predictive decision-making Simulation-based validation; limited real-world implementation [20] Stacking Ensemble + LSTM/LSTM EncoderDecoder Effective intrusion detection and QoS prediction; tested in live 5G network; low prediction error Complex model training; may require extensive data preprocessing; real-time computational cost [21] Hybrid RF-GRUAttention Model Accurate 6G traffic prediction; improves resource allocation; outperforms baseline GRU/LSTM; proactive optimization Model complexity; integration in real-time 6G deployment may be challenging [22] Survey on AI in Wireless Communication Broad review of ML, DL, RL applications in spectrum, resource, and signal management; identifies open challenges No experimental results; general survey; lacks specific implementation insights [10] AIAS (Supervised + Reinforcement Learning) High prediction accuracy (91.3%); 92.5% spectrum utilization; interferencefree allocation; 50% latency reduction Mainly simulation-based; deep learning integration not explored; potential scalability issues in very dense networks III. OVERVIEW OF SPECTRUM PREDICTION AND ALLOCATION MODELS This section provides an overview of model families used for spectrum occupancy prediction and channel, power, and access opportunity allocation in IoT centred wireless systems. Organization is along a realistic pipeline. First, create solid short and medium horizon channel state and demand predictions. Next, translate those forecasts, with real-time sensing, into allocation decisions that meet coexistence and latency requirements. A. Computational method classification Traditional machine learning Feature engineered classifiers remain a solid baseline for occupancy detection and short horizon prediction. All of support vector machines, random forests, gradient boosted trees, and k nearest neighbors operate on statistics such as received power percentiles, duty cycle, kurtosis, cyclostationary descriptors, and compact spectral aggregates. On real captures in sub six gigahertz bands, random forest and gradient boosting achieve high accuracy with moderate compute and clear feature importance, which is preferable in the context of embedded gateways and swift deployment [23]. Unsupervised learning also gives a label free baseline in cooperative sensing. Local decisions are improved by K means clustering of local decisions to make robustness to fading better than logical fusion rules under mobility [24]. On TV bands and wideband measurements, conventional pipelines that combine SVM or decision trees with dimensionality reduction offer competitive precision with lightweight inference and training [25]. In low signal to noise ratio narrowband cases, good descriptors in combination with SVM optimize robustness without keeping memory footprints small. [26] Measurement campaigns across hundreds of megahertz illustrate that machine learning is able to identify white space patterns recursing over time and location yet remain site specific to calibrate for maximum performance [27]. Regression style formulations of sensing wherein a model predicts probability of occupancy or duty cycle help downstream policies reason about risk instead of binary conditions. Experimental experiments illustrate robust calibration using traditional regressors on ! ! !!!!!!!!! ! 652 results are rarely reported with shared preprocessing, exact train/validation/test splits. The gap bars progress cumulatively and hide which gains persist robustly across environment-specific ones VI. FUTURE RESEARCH DIRECTIONS Several directions have the potential to advance the field meaningfully while being deployment constraints aware. Data and benchmarks need to first evolve from single-shot snapshots to multi-site, multi-band corpora cared for by communities with standardized divisions, measurement manifestos, and stable baseline code. Weakand semi-supervised labeling and synthetic-to-real domain randomization can scale sparse annotations without sacrificing validity. Privacy-preserving sharing—federated aggregation of feature stats or model updates—needs to be integrated into dataset programs day one. Second, RF foundation models deserve rigorous exploration. Self-supervised I/Q and spectrogram encoders over large corpora (masked modeling, contrastive learning) can provide transferable occupancy, device ID, and anomaly representations. Embedding graph inductive biases and topology tokens in spatiotemporal transformers has the potential to unleash generalization across bands and cities, especially when paired with lightweight, uncertainty-aware heads. Third, uncertainty and safety by design need to move from add-ons to first-class objectives. Predictors must report calibrated distributions (or conformal prediction sets) to allow allocators to optimize risk-sensitive measures (e.g., CVaR) rather than mere expected throughput. For control, reinforcement learning can leverage logged interactions for offline and safe learning, and model-based RL and shielded policies reduce trial-and-error in the field. For graphstructured optimization, unfolded solvers with learned components remain a high-leverage path to fixedlatency decisions with tight guarantees. Fourth, efficiency and sustainability must be built in. Distillation, pruning, low-rank adaptation, and quantization must be the default; event-driven and sparse operators can limit multiply-accumulate operations on edge hardware. Split learning across device–gateway–cloud stacks can allow compute to keep up with privacy and latency budgets. Fifth, continuous and federated learning should make itself deployable rather than experimental. Adaptive aggregation, heterogeneity-aware client selection, sitespecific adapters, and secure protocols can keep models current in non-IID drift with bounded backhaul. Finally, ecosystem integration needs to be handled. Open RAN-type controllers (near-RT RIC) will naturally find homes for spectrum-aware xApps; standardized APIs, policy schemas, and telemetry formats will reduce bespoke glue code. An effective MLOps for RF toolchain—dataset/version registries, drift alarms, reproducibility checks, and canary rollouts, will accelerate safe iteration. VII. CONCLUSION This research framed AI-driven dynamic spectrum prediction and assignment as a feasible pipeline for IoT networks: predict short-horizon usage and interference with calibrated models; allocate channels and power with structured controllers (unfolded optimizers and graph networks) or adaptive policies (reinforcement learning) with safety shields around them; and validate through closed-loop experiments with reproducible protocols. The review compared computational paradigms, observed benchmarks and datasets, and gleaned engineering recommendations for latencyand energy-constrained deployments. Python-centric implementation sections condensed these concepts into concrete tooling, data manifests, optimized baselines, sequence and vision encoders, graph-based allocators, RLlib training, ns-3 evaluation, edge deployment with quantization and federated updates. The future increasingly hinges ! ! !!!!!!!!! ! 653 on tightly controlled, uncertainty-aware, safety-focused designs that generalize across sites, bands, and time, rather than continuing to create ever-heavier models. Open-preprocessing community benchmarks with risk-sensitive performance metrics and closed-loop testing on standardized platforms will winnow out what works. From those, spectrum-intelligent IoT networks can mature from promising proof-ofconcept prototypes into robust, scalable systems that share sensibly and efficiently in shared radio environments. REFERENCES 1. Xing, X., Jing, T., Cheng, W., Huo, Y., & Cheng, X. (2013). Spectrum prediction in cognitive radio networks. IEEE Wireless Communications, 20(2), 90–96. https://doi.org/10.1109/MWC.2013.6507392 2. Pandian, P., Selvaraj, C., Natarajan, B., D. K. G., A., & Saikrishnan, S. (2023). Machine learning based spectrum prediction in cognitive radio networks. Proceedings of the IEEE International Conference on Networking and Communications (ICNWC), 1–6. https://doi.org/10.1109/ICNWC57852.2023.10127512 3. Hameed, A., Abbas, S. A., Nawaz, T., & Jumani, A. (2022). A deep learning approach for IoT traffic multi-classification in a smart-city scenario. IEEE Access, 10, 21193–21211. https://doi.org/10.1109/ACCESS.2022.3152600 4. Mosleh, S., Ma, Y., Rezac, J. D., & Coder, J. B. (2020). Dynamic spectrum access with reinforcement learning for unlicensed access in 5G and beyond. Proceedings of the IEEE 91st Vehicular Technology Conference (VTC2020-Spring), 1–6. 5. Aygül, M. A., Nazzal, M., Ekti, A. R., Görçin, A., da Costa, D. B., Ateş, H. F., & Arslan, H. (2020). Efficient spectrum occupancy prediction exploiting multidimensional correlations through composite 2D-LSTM models. Sensors, 21(1), 135. https://doi.org/10.3390/s21010135 6. Joneidi, M., Alkhouri, I., & Rahnavard, N. (2019). Large-scale spectrum occupancy learning via tensor decomposition and LSTM networks. arXiv preprint arXiv:1905.04392. https://doi.org/10.48550/arXiv.1905.04392 7. Pan, G., Li, J., & Li, M. (2025). Multi-channel multi-step spectrum prediction using Transformer and stacked Bi-LSTM. arXiv preprint arXiv:2405.19138. https://doi.org/10.48550/arXiv.2405.19138 8. Sun, Q., Fang, Z., Li, Y., & Petrosian, O. (2025). Jumping-knowledge graph attention networks for resource allocation in wireless networks. Scientific Reports, 15(1), 17459. https://doi.org/10.1038/s41598-025-84262-1 9. Abdisarabshali, P., Accurso, N., Malandra, F., Su, W., & Hosseinalipour, S. (2023). Synergies between federated learning and O-RAN: Towards an elastic virtualized architecture for multiple distributed machine learning services. arXiv preprint arXiv:2305.02109. 10. Dong, L., Qian, Y., & Xing, Y. (2022). Dynamic spectrum access and sharing through actor–critic deep reinforcement learning. EURASIP Journal on Wireless Communications and Networking, 2022(48). https://doi.org/10.1186/s13638-02202124-4 11. Alahakoon, C. A. D., Kandeepan, S., Yu, X., & Baldini, G. (2025). LSTM-based spectrum occupancy prediction performance with thin plate spline model. Proceedings of the IEEE International Conference on Information Networking (ICOIN). https://doi.org/10.1109/ICOIN63865.2025.10992844 12. Chang, H.-H., Song, H., Yi, Y., Zhang, J. (C.), He, H., & Liu, L. (2018). Distributive dynamic spectrum access through deep reinforcement learning: A reservoir computing based approach. arXiv preprint arXiv:1810.11758. 13. Bharathi, P. S., & Shankar, A. (2025). AIAS: Experimental evaluation of spectrum utilization over cognitive radio networks using artificial intelligence (AI) assisted adaptive searching algorithm. Proceedings of the Fifth International Conference on Advances in Electrical, Computing, Communication and Sustainable Technologies (ICAECT), 1–6. 14. Sekaran, R., Goddumarri, S. N., Kallam, S., Ramachandran, M., Patan, R., & Gupta, D. (2021). 5G integrated spectrum selection and spectrum access using AI-based framework for IoT-based sensor networks. Computer Networks, 186, 107649. 15. Ravi, B., & Verma, U. (2025). Spectrum allocation in 5G and beyond intelligent ubiquitous networks. International Journal of Network Management, 35(1), e2315. 16. Lin, K., Li, C., Tian, D., Ghoneim, A., Hossain, M. S., & Amin, S. U. (2019). Artificial-intelligence-based data analytics for cognitive communication in heterogeneous wireless networks. IEEE Wireless Communications, 26(3), 83–89. ! ! !!!!!!!!! ! 654 17. Al-Fatlawy, R. R., Rao, T. P., Bhariya, O. P., Gupta, S. K., & Rana, J. M. (2024). AI-enabled dynamic spectrum sharing in next-generation networks. Proceedings of the IEEE International Conference on Communication, Computing and Signal Processing (IICCCS), 1–6. 18. Alkurd, R., Abualhaol, I. Y., & Yanikomeroglu, H. (2020). Personalized resource allocation in wireless networks: An AIenabled and big data-driven multi-objective optimization. IEEE Access, 8, 144592–144609. 19. Mohammed, S. I. (2022). Exploring artificial intelligence (AI) techniques for forecasting network traffic: Network QoS and security perspectives (Master’s thesis). The University of Western Ontario, Canada. 20. Anis, M., Shah, A. F. M. S., Islam, M. B., Ekoru, J., & Madahana, M. (2025). Hybrid model for 6G network traffic prediction and wireless resource optimisation. IEEE Access. 21. Karthick, R. (2025). Artificial intelligence in wireless communication: Trends, applications, and future directions. 22. Reddy, S. V., Venkateswarlu, Y., Kalluri, P. R., & Reddy, M. S. (2024). Performance analysis of machine learning models on sub six gigahertz bands. Cluster Computing. 23. Abusubaih, M. A., & Khamayseh, S. (2022). Performance of machine learning based techniques for spectrum sensing in mobile cognitive radio networks. IEEE Access. 24. Mohammad, A., Awin, F., & Abdel Raheem, E. (2022). Case study of television spectrum sensing model based on machine learning techniques. Ain Shams Engineering Journal, 13(2). 25. Sheng, J., Meng, Y., Li, D., & Zou, Y. (2022). TCVQ SVM algorithm for narrowband spectrum sensing. Physical Communication, 50. 26. Alonso, H., Ponce, J. L., Valiente, D., & Navas, R. (2023). White spaces pattern finding and inference based on machine learning for multi frequency spectrum footprints. Computer Networks. 27. Begum, A. R., Divya Rani, M. S., Reddy, V. V., Reddy, B. S. K., & Papachari, B. (2024). Experimental testing and analysis of regression algorithms for spectrum sensing in cognitive radio networks. Wireless Networks. 28. Alghamdi, A. K., & Alsamhi, S. H. (2020). Prediction of radio frequency spectrum occupancy using support vector regression on cumulant features. Proceedings of the IEEE TrustCom. 29. Scalingi, A., Giustiniano, D., Calvo Palomino, R., Apostolakis, N., & Bovet, G. (2023). ElectroSense PSD spectrum dataset. Zenodo. https://doi.org/10.5281/zenodo.7521246 30. Yapar, C., Levie, R., Kutyniok, G., & Caire, G. (2022). Dataset of path loss and time of arrival radio maps with localization application. arXiv preprint arXiv:2212.11777. 31. Yang, K., Ding, Z., Fan, J., & Yang, Y. (2024). DC Net: A distant range content interaction network for radio map estimation. Digital Signal Processing. 32. Zhang, Y., Nie, Z., Han, P., & Liu, L. (2025). RMTransformer: Accurate radio map construction and coverage prediction. arXiv preprint arXiv:2501.05190. 33. Zhang, Y., Nie, Z., Han, P., & Liu, L. (2025). RadioFormer: A multiple granularity transformer for fast radio map estimation. arXiv preprint arXiv:2504.19161. 34. Zhang, Y., Botelho, P., Gordon, T., Zussman, G., & Kadota, I. (2025). Fair dynamic spectrum access via fully decentralized multi-agent reinforcement learning. arXiv preprint arXiv:2503.24296. 35. Ukpong, U. C., Idowu Bismark, O. A., Adetiba, E., & Kala, J. R. (2024). Deep reinforcement learning agents for dynamic spectrum access in television whitespace cognitive radio networks. Array. 36. Chowdhury, A., Verma, G., Rao, C., Swami, A., & Segarra, S. (2021). Unfolding WMMSE using graph neural networks for efficient power allocation. IEEE Transactions on Wireless Communications, 20(9), 6004–6017. 37. RadioMap3DSeer Project. (2024–2025). The First Pathloss Radio Map Prediction Challenge: Dataset and tracks. 38. Dhore, V., Bhat, A., Nerlekar, V., Chavhan, K., & Umare, A. (2024). Enhancing explainable AI by combining Grad-CAM and LRP for CNN interpretability. arXiv preprint arXiv:2405.12175. 39. Lu, W., Herman, D., Sun, G., Wang, Y., & Jaakkola, T. (2024). Causal state distillation for explainable reinforcement learning. Proceedings of the International Conference on Machine Learning. 40. Polese, M., D’Oro, S., Restuccia, F., Melodia, T., & Zorzi, M. (2025). From DeepSense to Open RAN: AI and ML advances in dynamic spectrum and RAN intelligence. arXiv preprint arXiv:2502.02889. 41. DeepSig. (2018). RadioML 2018.01A dataset. https://www.deepsig.io/datasets 42. RF Jamming Dataset. (2020). Real spectral scan measurements for wireless systems. https://example.com/rf-jammingdataset 43. Lyons, R. F. (2016). RML2016.10b: GNU radio simulated AMC dataset. https://www.deepsig.io/datasets 44. DeepRadar. (2022). DeepRadar2022 simulated radar signals dataset. https://example.com/deepradar2022 ! ! !!!!!!!!! ! 655 45. MIMO Radar Signal Dataset. (2021). Real array measurements with masked antennas. https://example.com/mimo-radardataset 46. WiSig. (2021). Wi-Fi fingerprinting dataset (Full WiSig). https://www.wisig.io 47. Pedregosa, F., et al. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825– 2830. 48. Paszke, A., et al. (2019). PyTorch: An imperative style, high-performance deep learning library. Proceedings of the NeurIPS Conference, 8024–8035. 49. Fey, M., & Lenssen, J. E. (2019). Fast graph representation learning with PyTorch Geometric. arXiv preprint arXiv:1903.02428. 50. Liang, E., et al. (2018). RLlib: Abstractions for distributed reinforcement learning. Proceedings of the International Conference on Machine Learning (ICML). 51. Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. (2019). Optuna: A next-generation hyperparameter optimization framework. Proceedings of the KDD Conference. 52. Gawłowicz, P., & Zubow, A. (2019). ns-3 meets OpenAI Gym: The playground for machine learning in networking research. Proceedings of the MSWiM Conference. 53. Henderson, T. R., Lacage, M., Riley, G. F., Dowell, C., & Kopena, J. (2008). Network simulations with the ns-3 simulator. Proceedings of the SIGCOMM Conference. 54. Beutel, D. J., et al. (2020). Flower: A friendly federated learning research framework. arXiv preprint arXiv:2007.14390.