Paper deep dive
TD-STGT: A Spatio-Temporal Graph Transformer for Mobile Traffic Demand Forecasting
Mohamad Alkadamani, Halim Yanikomeroglu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 9/12/2026, 2:51:09 AM
Summary
The paper introduces TD-STGT, a spatio-temporal graph transformer framework for forecasting fine-grained mobile traffic demand changes across geographic grids. It utilizes a population-scaled demand proxy derived from crowdsourced mobile measurements and daytime population data. Evaluated across five Canadian metropolitan regions, TD-STGT outperforms baseline models (including STGCN, LSTM, and XGBoost) in predicting grid-level demand changes, achieving a Delta R-squared of 0.462 and a 5.7% reduction in Delta RMSE. The model combines local graph attention, global spatial self-attention, temporal convolution, and temporal self-attention to capture complex spatio-temporal dependencies.
Entities (10)
Relation Signals (7)
TD-STGT → achieves → Delta R-squared
confidence 95% · reaching a ΔR2 of 0.462 and reducing ΔRMSE by 5.7% relative to the strongest baseline.
TD-STGT → evaluatedin → Canadian Metropolitan Regions
confidence 95% · Experiments across five Canadian metropolitan regions show that TD-STGT achieves the best performance
TD-STGT → uses → Crowdsourced Mobile Measurements
confidence 95% · The framework uses a population-scaled demand proxy developed from crowdsourced mobile measurements and daytime population information.
TD-STGT → outperforms → STGCN
confidence 92% · The strongest non-proposed baseline is STGCN with ΔRMSE of 0.053, while TD-STGT reduces ΔRMSE to 0.050, a 5.7% reduction.
TD-STGT → contains → Graph Attention Network (GAT)
confidence 90% · The proposed model... uses local graph attention, global spatial self-attention, temporal convolution, and temporal self-attention
TD-STGT → contains → Temporal Convolutional Network (TCN)
confidence 90% · It combines local graph attention network (GAT), global spatial attention, temporal convolutional network (TCN), and temporal transformer blocks.
Population-Scaled Demand Proxy → derivedfrom → Crowdsourced Mobile Measurements
confidence 90% · The framework uses a population-scaled demand proxy developed from crowdsourced mobile measurements and daytime population information.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Fine-grained mobile traffic demand forecasting is essential for long-term planning of 5G and future 6G networks, including radio upgrades, site densification, backhaul expansion, and spectrum activation. This paper proposes the Traffic Demand Spatio-Temporal Graph Transformer (TD-STGT), a graph neural forecasting framework for predicting changes in wireless mobile traffic demand across fine geographic grids. The framework uses a population-scaled demand proxy developed from crowdsourced mobile measurements and daytime population information. Experiments across five Canadian metropolitan regions show that TD-STGT achieves the best performance in forecasting grid-level demand changes, reaching a $\Delta R^2$ of 0.462 and reducing $\Delta$RMSE by 5.7\% relative to the strongest baseline. The proposed model provides a practical tool for identifying areas with increasing demand pressure and prioritizing future mobile-network capacity upgrades.
Tags
Links
- Source: https://arxiv.org/abs/2609.06636v1
- Canonical: https://arxiv.org/abs/2609.06636v1
Trouble viewing inline? Open PDF directly →
Full Text
28,126 characters extracted from source content.
Expand or collapse full text
TD-STGT: A Spatio-Temporal Graph Transformer for Mobile Traffic Demand Forecasting Mohamad Alkadamani12 and Halim Yanikomeroglu2 Affiliation: 1Communications Research Centre Canada, Ottawa, ON, Canada Affiliation: 2Department of Systems and Computer Engineering, Carleton University, Ottawa, ON, Canada Affiliation: Email: mohamad.alkadamani@ised-isde.gc.ca, halim@sce.carleton.ca Abstract Fine-grained mobile traffic demand forecasting is essential for long-term planning of 5G and future 6G networks, including radio upgrades, site densification, backhaul expansion, and spectrum activation. This paper proposes the Traffic Demand Spatio-Temporal Graph Transformer (TD-STGT), a graph neural forecasting framework for predicting changes in wireless mobile traffic demand across fine geographic grids. The framework uses a population-scaled demand proxy developed from crowdsourced mobile measurements and daytime population information. Experiments across five Canadian metropolitan regions show that TD-STGT achieves the best performance in forecasting grid-level demand changes, reaching a ΔR2 R^2 of 0.462 and reducing Δ by 5.7% relative to the strongest baseline. The proposed model provides a practical tool for identifying areas with increasing demand pressure and prioritizing future mobile-network capacity upgrades. Index Terms: mobile traffic forecasting, spatio-temporal graph neural networks, graph transformers, network planning. I Introduction Mobile network traffic demand continues to increase as LTE and 5G networks support high-resolution video, cloud services, fixed-wireless access, connected vehicles, and emerging 6G use cases [1]. For mobile network operators (MNOs), anticipating where demand will grow is essential for planning decisions such as radio upgrades, site densification, backhaul expansion, and spectrum activation. A key challenge is that direct operator traffic data are proprietary and are generally not accessible for research. This study therefore uses crowdsourced mobile measurement data as an observable basis for traffic-demand analysis. In this context, crowdsourced data refers to time-stamped and geo-located user-side measurements collected from mobile devices through software development kits (SDKs) embedded in applications. Such data do not provide the complete carried traffic of an operator network, but they provide repeated observations of mobile-network activity over space and time. Prior spectrum-demand studies used geospatial features, deployment-based proxies, interpretable learning, and graph models to estimate local demand [2, 3, 4, 5, 6], but they mainly address spatial estimation. Mobile traffic forecasting has used recurrent and graph-based models to capture temporal and spatial dependence [7, 8, 9, 10]; however, many studies rely on proprietary counters or short observation windows. This paper proposes a data-driven framework that constructs a population-scaled proxy from crowdsourced measurements and forecasts grid-level demand change using a spatio-temporal graph transformer. Spatio-temporal modeling is used because demand does not evolve independently across grids: nearby and functionally related areas often change together. The proposed model, referred to as Traffic Demand Spatio-Temporal Graph Transformer (TD-STGT), uses local graph attention, global spatial self-attention, temporal convolution, and temporal self-attention to forecast grid-level demand changes. The contributions are summarized as follows: • A crowdsourced proxy for mobile traffic demand is developed from user-side mobile measurement data. • A spatio-temporal graph transformer architecture is formulated for grid-level mobile traffic-demand forecasting. • An assessment is conducted across five Canadian metropolitan regions, demonstrating the ability of the proposed model to forecast demand-change patterns in urban and suburban planning scenarios. The remainder of the paper is organized as follows. Section I presents the analysis setup and proxy construction. Section I formulates the forecasting problem and describes TD-STGT. Section IV presents the experimental setup and results. Section V concludes the paper. I Analysis Setup and Proxy Construction I-A Analysis Setup The study covers Montreal, Ottawa, Toronto, Calgary, and Vancouver, focusing on urban and suburban areas where mobile traffic growth is concentrated. Each region is divided into grid cells of approximately 1.5km21.5\,km^2, yielding 38,246 grids. Crowdsourced observations are spatially assigned to grids and aggregated weekly and monthly; monthly aggregation is used as the main planning-level granularity because it is more stable under sampling variability. The data include SDK-based crowdsourced measurements, daytime population, and network deployment records. The crowdsourced data provide user-side observations of mobile activity, daytime population approximates the number of potential users present in each grid during peak activity periods, and deployment records are used only for independent validation. I-B Population-Scaled Peak-Demand Proxy The proposed proxy preserves user-side information in transmitted-byte measurements while reducing bias from uneven SDK sampling and low-sample grids. Let ℳg,tpM_g,t^p be the set of crowdsourced observations in grid g during peak hours of period t, and let bmb_m be the transmitted bytes of observation m. The observed peak-hour byte volume and record count are Bg,tp=∑m∈ℳg,tpbm,Rg,tp=|ℳg,tp|.B^p_g,t= _m _g,t^pb_m, R^p_g,t=|M_g,t^p|. (1) The exposure-normalized peak intensity is βg,tp=Bg,tpRg,tp+ϵ, _g,t^p= B^p_g,tR^p_g,t+ε, (2) where ϵε avoids division by zero. To stabilize low-sample grids, the grid-specific value is shrunk toward the corresponding city-period average: β¯c,tp=∑g∈cBg,tp∑g∈cRg,tp+ϵ,wg,t=Rg,tpRg,tp+κ, β_c,t^p= _g∈ cB^p_g,t _g∈ cR^p_g,t+ε, w_g,t= R^p_g,tR^p_g,t+κ, (3) where c is the city containing grid g and κ controls shrinkage strength. The stabilized peak-hour intensity is β~g,tp=wg,tβg,tp+(1−wg,t)β¯c,tp. β_g,t^p=w_g,t _g,t^p+(1-w_g,t) β_c,t^p. (4) When a grid has many peak-hour samples, wg,tw_g,t approaches one; otherwise, the estimate is pulled toward the city-period average. Let PgdayP^day_g denote daytime population. The population-scaled proxy is Dg,tpop=Pgdayβ~g,tp.D^pop_g,t=P^day_g β_g,t^p. (5) This proxy represents grid-level peak-demand potential by combining observed mobile data usage per crowdsourced observation with the daytime population in each grid. The level series is log-transformed as yg,t=log(1+Dg,tpop)y_g,t= (1+D^pop_g,t), and the forecasting target is the one-period change Δyg,t=yg,t−yg,t−1 y_g,t=y_g,t-y_g,t-1. The proxy is computed at weekly and monthly resolution. Weekly aggregation provides more temporal samples, while monthly aggregation is used as the main planning-level target because it smooths sampling variability and aligns with MNO capacity-planning cycles. I-C Proxy Behavior and Validation Figure 1 summarizes the temporal behavior of the proxy before forecasting. Panel (a) indexes the city-level aggregate series and highlights broad relative growth patterns across metropolitan regions. Panel (b) shows representative high-demand Toronto grids and illustrates that even within one city, local trajectories can evolve differently. This heterogeneity motivates grid-level spatio-temporal forecasting rather than a purely aggregate city-level predictor. Fig. 1: Temporal evolution of the proposed population-scaled peak-demand proxy. In Fig. 2, each grid is colored by the annualized trend of its monthly log-proxy series. The spatially heterogeneous growth pattern in Toronto further motivates graph-based forecasting. Fig. 2: Spatial rate of change of the proposed proxy in Toronto. Network deployment records are used for independent validation. A meaningful demand proxy should exhibit positive spatial association with deployed bandwidth, because operators tend to provision greater capacity in areas with persistent demand pressure. For validation, a site-data snapshot is assigned to the same grid. Let BWg,tBW_g,t denote total deployed bandwidth in grid g [5]. The association is measured as ρt(Dpop,BW)=Covg(Dg,tpop,BWg,t)σD,tσBW,t, _t(D^pop,BW)= Cov_g(D^pop_g,t,BW_g,t) _D,t _BW,t, (6) where Covg(⋅)Cov_g(·) is covariance across grids and σD,t _D,t and σBW,t _BW,t are the corresponding standard deviations. I Problem Formulation The objective is to forecast changes in grid-level mobile traffic demand represented by the crowdsourced population-scaled proxy. Let t∈ℝNy_t ^N denote the log-transformed proxy values for all N grids at period t, where yg,t=log(1+Dg,tpop)y_g,t= (1+D^pop_g,t). Because direct level forecasting is highly persistent, the main supervised target is the one-period change: Δt=t−t−1. _t=y_t-y_t-1. (7) This formulation evaluates whether a model can identify grids whose demand trajectory changes beyond simple persistence. Let t∈ℝN×FX_t ^N× F denote available period-level features, including historical proxy values, previous changes, lagged peak-hour sample counts, exposure-normalized byte indicators and daytime population descriptors. The grid-time panel contains N=10,250N=10,250 grids across the five metropolitan regions. For an input window of length L, the model receives historical proxy states, change values, and auxiliary grid-period features from t−Lt-L to t−1t-1 and predicts future proxy changes as Δ^t:t+H−1=fΘ(t−L:t−1,), _t:t+H-1=f_ (X_t-L:t-1,G ), (8) where t−L:t−1X_t-L:t-1 is the spatio-temporal input tensor, G is the grid graph, H is the forecast horizon, and Θ denotes trainable parameters. The main task uses H=1H=1, and evaluation reports ΔR2 R^2, Δ , and Δ on the change target. IV Proposed TD-STGT Methodology IV-A Graph Construction A spatial graph =(,ℰ)G=(V,E) is constructed where each node represents a grid cell. Edges encode spatial adjacency and distance-based proximity. The edge weight between nodes i and j is Aij=exp(−‖si−sj‖22σ2),j∈k(i),0,otherwise,A_ij= cases (- \|s_i-s_j\|_2^2σ^2 ),&j _k(i),\\ 0,&otherwise, cases (9) where sis_i and sjs_j are grid centroids, k(i)N_k(i) contains the k nearest grids, and σ controls distance decay. Self-loops preserve each grid’s history; the default setting uses k=8k=8. IV-B Overall Architecture TD-STGT contains spatial, temporal, and prediction modules. It combines local graph attention network (GAT), global spatial attention, temporal convolutional network (TCN), and temporal transformer blocks. Fig. 3 summarizes the architecture. Fig. 3: Overview of the TD-STGT architecture. IV-C Input Embedding The input window is projected into a d-dimensional representation: (0)=ϕe(t−L:t−1)=t−L:t−1e+e,H^(0)= _e(X_t-L:t-1)=X_t-L:t-1W_e+b_e, (10) where (0)∈ℝN×L×dH^(0) ^N× L× d. IV-D Local Spatial Graph Attention Branch For each time step τ, a graph attention layer updates node i using its spatial neighbors: eijτ=LeakyReLU(⊤[Gi,τ∥Gj,τ]),e_ij^τ=LeakyReLU (a [W_Gh_i,τ\|W_Gh_j,τ] ), (11) αijτ=exp(eijτ)∑k∈(i)exp(eikτ), _ij^τ= (e_ij^τ) _k (i) (e_ik^τ), (12) i,τG=σ(∑j∈(i)αijτGj,τ),z^G_i,τ=σ ( _j (i) _ij^τW_Gh_j,τ ), (13) where eijτe_ij^τ is the unnormalized attention score from neighbor j to grid i, αijτ _ij^τ is the normalized attention coefficient, a is a learnable attention vector, GW_G is a graph-attention projection matrix, i,τh_i,τ is the input embedding of grid i at time τ, ∥\| denotes concatenation, and σ(⋅)σ(·) is a nonlinear activation. Multi-head attention is implemented by concatenating or averaging head-specific outputs. IV-E Global Spatial Transformer Branch The spatial transformer captures long-range relationships using self-attention. Spatial positional encodings are added as ~:,τ,:=(0):,τ,:+S. H_:,τ,:=H^(0)_:,τ,:+P^S. (14) For each time step, multi-head spatial self-attention is computed as MSSA(~)=Concat(head1,…,headh)O,MSSA( H)=Concat(head_1,…,head_h)W^O, (15) headm=softmax(mS(mS)⊤dh)mS,head_m=softmax ( Q_m^S(K_m^S) d_h )V_m^S, (16) where h is the number of heads, dhd_h is the per-head dimension, and mSQ_m^S, mSK_m^S, and mSV_m^S are query, key, and value matrices, respectively. The branch output is S=LN(MSSA(~)+~),Z^S=LN (MSSA( H)+ H ), (17) where LN(⋅)LN(·) denotes layer normalization. IV-F Gated Spatial Fusion The local GAT output GZ^G and global spatial transformer output SZ^S are fused through a nonlinear gate: S=sigmoid(Gg1+Sg2+g),G^S=sigmoid (Z^GW_g1+Z^SW_g2+b_g ), (18) SG=S⊙S+(1−S)⊙G,Z^SG=G^S ^S+(1-G^S) ^G, (19) where SG^S is the learned fusion gate, g1W_g1 and g2W_g2 are trainable projection matrices, gb_g is a bias term, and ⊙ denotes element-wise multiplication. The gate allows each grid and timestamp to determine whether local or global spatial context is more informative. IV-G Temporal Dependency Modeling The temporal module begins with a TCN applied to each grid sequence, using causal dilated convolutions to model local temporal dependencies. A causal dilated convolution is i,τ=∑q=0K−1qi,τ−q⋅drSG,o_i,τ= _q=0^K-1W_qZ^SG_i,τ-q· d_r, (20) where i,τo_i,τ is the TCN output for grid i at time τ, K is the convolution kernel size, drd_r is the dilation rate, and qW_q is the convolution weight at lag q. Residual TCN blocks capture local temporal changes in deployment and demand. Temporal positional encodings are then added: ~=+SG+T, Z=O+Z^SG+P^T, (21) where O is the TCN representation and TP^T encodes the order of the input periods. Temporal multi-head self-attention is then applied along the time dimension for each grid: MTSA(~)=Concat(thead1,…,theadh)T,O,MTSA( Z)=Concat(thead_1,…,thead_h)W^T,O, (22) theadm=softmax(mT(mT)⊤dh)mT,thead_m=softmax ( Q_m^T(K_m^T) d_h )V_m^T, (23) where mTQ_m^T, mTK_m^T, and mTV_m^T are temporal query, key, and value matrices for head m, and T,OW^T,O is the temporal output projection. The temporal representation is T=LN(FFN(LN(MTSA(~)+~))+~),Z^T=LN (FFN (LN(MTSA( Z)+ Z) )+ Z ), (24) where FFN(⋅)FFN(·) is a position-wise feed-forward network. IV-H Prediction Head and Loss Function The prediction head maps the temporal representation to the forecast horizon of proxy changes: Δ^t:t+H−1=ϕout(T), _t:t+H-1= _out(Z^T), (25) where ϕout(⋅) _out(·) is implemented using two 1×11× 1 convolutional layers with a ReLU activation followed by a fully connected projection to the forecast horizon. The model is trained using a loss function that combines change-forecasting error, spatial smoothness, and temporal smoothness: ℒ=ℒΔ+λsℒS+λtℒT,L=L_ + _sL_S+ _tL_T, (26) where λs _s and λt _t control spatial and temporal regularization: ℒΔ=1NH∑g=1N∑h=0H−1(Δyg,t+h−Δy^g,t+h)2,L_ = 1NH _g=1^N _h=0^H-1 ( y_g,t+h- y_g,t+h )^2, (27) ℒS=1|ℰ|H∑(i,j)∈ℰ∑h=0H−1(Δy^i,t+h−Δy^j,t+h)2,andL_S= 1|E|H _(i,j) _h=0^H-1 ( y_i,t+h- y_j,t+h )^2,and (28) ℒT=1N(H−1)∑g=1N∑h=1H−1(Δy^g,t+h−Δy^g,t+h−1)2,L_T= 1N(H-1) _g=1^N _h=1^H-1 ( y_g,t+h- y_g,t+h-1 )^2, (29) where ℒΔL_ penalizes change-forecasting error, ℒSL_S penalizes unrealistic discontinuities in predicted changes between connected grids, and ℒTL_T regularizes abrupt oscillations across multi-step change forecasts. V Experimental Setup, Results, and Discussion V-A Dataset Split and Sliding Window Strategy Chronological splitting is used to avoid future information leakage: earlier periods are used for training, an intermediate period for validation, and the most recent periods for testing. Validation and test samples use only historical context from preceding periods, so no future proxy value or current-period proxy component enters the input. To assess spatial generalization, an additional blocked split holds out contiguous grid regions during training. The strictest protocol combines both dimensions by testing future periods in spatially held-out regions. The main results report the chronological test setting, while the spatial and spatio-temporal splits are used for robustness analysis. The main configuration uses an input window of approximately four years and H=1H=1 forecast step, corresponding to next-period proxy-change forecasting at the selected weekly or monthly granularity. V-B Model Configuration Table I lists the default TD-STGT configuration. All models are implemented in PyTorch 2.7.1 and trained on an AWS EC2 G5.2xlarge instance equipped with one NVIDIA A10G GPU with 24 GB GPU memory. Hyperparameters are selected through validation-set grid search over reasonable ranges commonly used for spatio-temporal forecasting. TABLE I: TD-STGT model configuration Parameter Value Grid nodes N 38,246 Input features F 18 proxy-derived and auxiliary features Input window L 48 snapshots Forecast horizon H 1 snapshot Graph neighbors k 8 Embedding dimension d 128 Attention heads h 4 TCN kernel size K 3 Dropout 0.20 Optimizer AdamW Learning rate 1×10−31× 10^-3 Batch size 32 windows Maximum epochs 100 Early stopping 12 epochs AWS instance EC2 G5.2xlarge GPU NVIDIA A10G 24 GB V-C Baseline Models and Evaluation Metrics Seven baseline models are selected: linear autoregressive regression, RF, XGBoost, LSTM, GRU, CNN-LSTM, and STGCN. The linear model provides an interpretable time-series baseline. RF and XGBoost evaluate strong nonlinear tabular approaches. LSTM and GRU evaluate sequence models without explicit spatial graph structure. CNN-LSTM evaluates Euclidean spatio-temporal learning. STGCN evaluates graph-based spatio-temporal forecasting. Table I summarizes the main hyperparameters used for the baseline models. TABLE I: Key hyperparameters of baseline models Model Hyperparameter Value Linear ARX Regularization Ridge, α=1.0α=1.0 RF Estimators, depth 300, 18 XGBoost Estimators, LR, depth 500, 0.03, 6 LSTM Layers, hidden, dropout 2, 96, 0.2 GRU Layers, hidden, dropout 2, 96, 0.2 CNN-LSTM Kernel, hidden, dropout 3, 96, 0.2 STGCN Layers, channels, kernel 2, 64, 3 TD-STGT Heads, embedding, dropout 4, 128, 0.2 Performance is evaluated with ΔR2 R^2, Δ , and Δ on the one-period change target. These metrics avoid overstating skill from the high persistence of proxy levels. V-D Proxy Validation Results Table I reports deployed-bandwidth validation results for alternative proxy variants. The proposed population-scaled peak-demand proxy reaches the strongest association with deployed bandwidth, with median Pearson correlation of 0.71 and OLS R2R^2 of 0.66. TABLE I: Proxy validation against deployed bandwidth Proxy variant Median ρ OLS R2R^2 p-value Observed all-day bytes/hour 0.64 0.60 <0.001<0.001 Population-scaled peak-demand proxy 0.71 0.66 <0.001<0.001 V-E Forecasting Performance Table IV compares TD-STGT against the persistence and learning baselines for one-step-ahead change forecasting. The no-change baseline predicts Δyg,t=0 y_g,t=0 and is included because the proxy level is highly persistent. TD-STGT achieves the highest ΔR2 R^2 and the lowest Δ and Δ . The strongest non-proposed baseline is STGCN with Δ of 0.053, while TD-STGT reduces Δ to 0.050, a 5.7% reduction. TABLE IV: One-step demand-change forecasting performance on the 2025 test set Model ΔR2↑ R^2 Δ ↓ Δ ↓ No-change baseline 0.000 0.069 0.047 Linear ARX 0.112 0.065 0.045 RF 0.246 0.060 0.041 XGBoost 0.312 0.057 0.039 LSTM 0.274 0.059 0.040 GRU 0.291 0.058 0.039 CNN-LSTM 0.341 0.056 0.038 STGCN 0.405 0.053 0.036 TD-STGT (Proposed) 0.462 0.050 0.034 The no-change baseline captures temporal persistence, while RF, XGBoost, and recurrent models provide moderate gains. STGCN is the strongest baseline because it uses graph structure, and TD-STGT performs best by combining graph locality, global spatial context, and temporal attention. Fig. 4 shows the distribution of absolute delta errors. TD-STGT has the lowest median error and one of the tightest spreads, while STGCN is the strongest baseline and the recurrent models show larger dispersion. Fig. 4: Absolute delta-error distributions across models. The robustness analysis trains each neural model under 20 random seeds. Fig. 5 shows the resulting Δ and Δ distributions. LSTM, GRU, and CNN-LSTM exhibit wider spreads, STGCN is more stable, and TD-STGT has the narrowest distribution, indicating the most consistent convergence. Fig. 5: Change-forecasting robustness across 20 random seeds. V-F Sensitivity Analysis Sensitivity is evaluated by varying learning rate, attention heads, batch size, and embedding dimension. TABLE V: Sensitivity analysis for key hyperparameters Hyperparameter Setting Δ Δ Learning rate 5×10−45× 10^-4 0.053 0.036 1×10−31× 10^-3 0.050 0.034 2×10−32× 10^-3 0.054 0.037 Attention heads 2 0.054 0.037 4 0.050 0.034 8 0.052 0.035 Batch size 16 0.053 0.036 32 0.050 0.034 64 0.052 0.035 Embedding dimension 64 0.056 0.038 128 0.050 0.034 256 0.050 0.034 As shown in Table V, the best operating point uses learning rate 10−310^-3, four heads, batch size 32, and embedding dimension 128. V-G Ablation Study The ablation study removes one TD-STGT component at a time under the same training protocol to identify where the gains originate. As shown in Table VI, removing the spatial module causes the largest degradation, confirming the importance of spatial context; proxy simplification also reduces accuracy. TABLE VI: Ablation analysis of TD-STGT Variant ΔR2↑ R^2 Δ ↓ Δ ↓ Full TD-STGT 0.462 0.050 0.034 Without spatial module 0.311 0.063 0.043 Without local GAT branch 0.421 0.052 0.036 Without spatial transformer 0.394 0.054 0.037 Without TCN 0.372 0.055 0.038 Without temporal transformer 0.386 0.054 0.037 Without user-weighted proxy mapping 0.335 0.061 0.041 V-H Model Complexity and Computational Cost Table VII compares advanced models in terms of parameter count, training time, inference time, and Δ . TD-STGT is the most expensive model, but the cost remains practical for offline planning and is justified by the accuracy gain. TABLE VII: Model complexity and computational cost Model Params Train Time Inference Δ (M) (min) (ms/batch) LSTM 0.31 38 1.7 0.059 GRU 0.27 36 1.5 0.058 CNN-LSTM 0.58 47 2.2 0.056 STGCN 1.12 59 2.9 0.053 TD-STGT 2.86 88 4.4 0.050 VI Conclusion This study presented a spatio-temporal framework for forecasting grid-level changes in mobile traffic demand using a crowdsourced population-scaled proxy. The framework converts user-side mobile measurements and daytime population information into a demand signal whose temporal changes can be modeled at fine spatial granularity. Pilot results show that TD-STGT achieves the best performance among the evaluated baselines, with ΔR2=0.462 R^2=0.462 and Δ of 0.050. The results indicate that proxy-driven change forecasting can help identify areas with increasing demand pressure and support proactive planning decisions such as densification, carrier activation, radio upgrades, and backhaul expansion. References [1] Ericsson (2025) Ericsson mobility report. Cited by: §I. [2] J. Parekh, A. Ghasemi, and H. Yanikomeroglu (2023) Data-driven modelling of mobile network demand for efficient spectrum management. In 2023 IEEE 34th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), p. 1–6. External Links: Document Cited by: §I. [3] J. Parekh, E. Yackoboski, A. Ghasemi, and H. Yanikomeroglu (2025) Modeling local demand for mobile spectrum: an interpretable machine learning approach. IEEE Open Journal of the Communications Society 6, p. 4063–4082. External Links: Document Cited by: §I. [4] J. Parekh, A. Ghasemi, and H. Yanikomeroglu (2024) DeepSpectrum: a localized demand estimation model for mobile spectrum using deep learning. In 2024 IEEE 100th Vehicular Technology Conference (VTC2024-Fall), Washington, DC, USA. Cited by: §I. [5] M. Alkadamani, A. Ghasemi, and H. Yanikomeroglu (2024) Towards flexible spectrum access: data-driven insights into spectrum demand. In 2024 IEEE 100th Vehicular Technology Conference (VTC2024-Fall), p. 1–7. Cited by: §I, §I-C. [6] M. Alkadamani, H. Yanikomeroglu, and A. Ghasemi (2025) A graph-based approach to spectrum demand prediction using hierarchical attention networks. In 2025 IEEE Global Communications Conference (GLOBECOM), Taipei, Taiwan. Cited by: §I. [7] C. Qiu, Y. Zhang, Z. Feng, P. Zhang, and S. Cui (2018) Spatio-temporal wireless traffic prediction with recurrent neural network. IEEE Wireless Communications Letters 7 (4), p. 554–557. External Links: Document Cited by: §I. [8] X. Wang, Z. Zhou, F. Xiao, K. Xing, Z. Yang, Y. Liu, and C. Peng (2019) Spatio-temporal analysis and prediction of cellular traffic in metropolis. IEEE Transactions on Mobile Computing 18 (9), p. 2190–2202. External Links: Document Cited by: §I. [9] X. Wang, K. Yang, Z. Wang, J. Feng, L. Zhu, J. Zhao, and C. Deng (2023) Adaptive hybrid spatial-temporal graph neural network for cellular traffic prediction. In IEEE International Conference on Communications (ICC), p. 4026–4032. Cited by: §I. [10] W. Jiang and J. Luo (2022) Graph neural network for traffic forecasting: a survey. Expert Systems with Applications 207. Cited by: §I.