user mobility modeling

This study addresses the limitations of existing mobile cellular traffic load prediction methods, which often underperform due to insufficient modeling of user dynamic behavior. Focusing on highway scenarios, this work proposes a machine learning–based joint prediction model that, for the first time, integrates crowd dynamics—capturing both user count and mobility—as a key input feature alongside historical traffic time-series data. Departing from prior approaches that rely primarily on increasingly complex models, the proposed method underscores the fundamental importance of high-quality, behaviorally informed data for enhancing prediction accuracy. Experimental results demonstrate that incorporating crowd dynamics alone improves prediction accuracy by approximately 60%, substantially outperforming baseline methods.

Modeling non-stationarity and abrupt changes in cellular base station traffic forecasting remains challenging due to the complex, time-varying nature of observational noise. Method: This paper introduces a novel paradigm centered on *structured noise priors*, revealing for the first time that mobile traffic noise exhibits learnable, decomposable dynamics. We propose NPDiff—a framework that explicitly disentangles noise into a dynamically aware prior component and an adaptive residual component—moving beyond conventional approaches that solely optimize denoising networks. The method integrates time-series modeling with a plug-and-play prior injection mechanism, ensuring compatibility with state-of-the-art diffusion-based predictors (e.g., CSDI, DLinear-Diff). Results: Evaluated on multiple real-world urban base station datasets, NPDiff achieves over 30% reduction in MAE and RMSE, while significantly improving robustness and inference efficiency—demonstrating the critical value of explicit noise prior modeling for edge-intelligent traffic forecasting.

Quantifying predictability in latency-sensitive communication networks remains an open challenge. Method: This paper formally defines system performance predictability and establishes a theoretical framework based on total variation distance. Predictability is measured as the total variation between the optimal predictive distribution and the marginal distribution. Modeling multi-hop networks under Markov dynamics, we integrate the Geo/Geo/1 queuing model with spectral analysis of Markov chains to derive exact and approximate closed-form predictability expressions—along with tight spectral upper bounds—for both single-hop and multi-hop scenarios. Contribution/Results: We quantitatively characterize how observation granularity and system dynamism jointly affect prediction capability. This work provides the first theoretical foundation for Quality-of-Service (QoS) proactive prediction and deterministic network adaptive design, enabling principled predictability-aware network optimization.

This work addresses the challenge of accurately modeling dynamic and spatially heterogeneous cellular traffic—particularly in emerging regions with scarce data—where conventional base station deployment and operation strategies fall short. We propose NetSpatial, the first system to integrate spatial conditional generative models into cellular traffic forecasting. By fusing multimodal urban context such as satellite imagery and points of interest, and employing a multi-level flow-matching architecture that decouples periodic patterns from residual dynamics, NetSpatial enables direct long-horizon traffic generation. The framework unifies “what-if” analysis for deployment planning and “what-to-do” optimization for operational decisions. Evaluated on real-world data, it reduces Jensen–Shannon divergence by 29.44%, demonstrates zero-shot generalization across cities, and achieves 16.8% energy savings through base station sleep scheduling and load balancing while maintaining quality of experience for over 80% of users.

This work addresses the challenge of generating human mobility trajectories that faithfully reproduce real-world network structures and temporal patterns without relying on assumptions about individual behavior. The authors propose a network-centric, privacy-preserving trajectory generation framework that constructs a time-varying Markovian dynamics model grounded in spatial interaction networks. The transition matrix is defined through a gravity-like distance decay function, exogenous temporal scheduling, and directional bias. Notably, the model introduces, for the first time, a periodic stationary population distribution as a non-transient reference state. By rigorously linking trajectory realizations to multi-step Markov dynamics, the method successfully reproduces structured origin–destination flows shaped by network geometry, temporal modulation, and connectivity constraints, achieving high consistency between individual-level trajectories and macroscopic dynamics, with discrepancies attributable solely to finite-population sampling effects.

This work addresses the limitation of existing mobile traffic forecasting models, which capture only static long-term temporal patterns and fail to model the complex interactions between traffic dynamics and adaptive network parameter adjustments. To overcome this, we propose MobiWM, a world model for mobile networks that treats traffic as a system state and explicitly models its dynamic evolution in response to network actions—such as transmit power, azimuth, and mechanical/electrical tilt—while integrating multimodal environmental context from both images and sequential data. MobiWM enables open-ended, continuous action trajectory rollouts over unlimited time horizons and, for the first time, introduces world models to mobile traffic extrapolation by constructing an explorable counterfactual simulation environment. Evaluated on real-world variable-parameter data spanning 9 regions and 31,900 cells, MobiWM significantly outperforms existing methods in distributional fidelity and effectively supports downstream reinforcement learning optimization, paving the way for digital twin–driven wireless network management.

This work addresses the limitation of existing traffic forecasting methods, which are predominantly deterministic and thus struggle to effectively capture the inherent uncertainty in traffic dynamics. To overcome this, the authors propose a general, plug-and-play probabilistic framework that transforms any pre-existing model into a probabilistic one by replacing its output layer with a Gaussian Mixture Model (GMM) layer and training end-to-end using negative log-likelihood loss—without altering the original model architecture or training pipeline. The approach enables, for the first time in traffic forecasting, a systematic evaluation of cumulative distribution functions and prediction intervals, substantially improving uncertainty quantification accuracy. Experiments on multiple real-world datasets demonstrate that the method preserves the original deterministic performance while outperforming both unimodal probabilistic and deterministic baselines, and exhibits enhanced robustness under low-quality data conditions.

This study addresses the limited generalizability of existing traffic prediction methods across diverse route choice scenarios, which fail to accurately capture network-wide travel time variations arising from differing path allocations under identical travel demand. To overcome this limitation, the paper proposes a Generalized Travel Time Predictor (GenTTP)—the first framework capable of generalizing travel time predictions across a wide spectrum of routing strategies. GenTTP leverages graph neural networks to jointly model spatiotemporal traffic dynamics and microscopic route choice behavior, effectively capturing complex interactions within the road network. Experimental results demonstrate that GenTTP significantly improves the accuracy of both travel time and traffic flow predictions under various path assignment scenarios, exhibiting strong generalization capabilities even for atypical and dynamically evolving travel behaviors.

This work addresses the challenge of achieving both high prediction accuracy and scalability in resource-constrained centralized Wi-Fi controllers, where a single global model struggles to meet the demands of large-scale networks. The authors propose a cluster-oriented, customized modeling paradigm that first groups Wi-Fi time-series data through feature engineering and principal component analysis (PCA), then constructs dedicated prediction models for each cluster. This approach significantly reduces the mean absolute error (MAE) while effectively balancing predictive accuracy against system resource consumption. By enabling selective deployment and adaptive network management, the method achieves a synergistic optimization of scalability and precision, making it well-suited for dynamic, large-scale Wi-Fi environments.