AdaST: Adaptive Coupling for Spatial-Temporal Forecasting¶
Quick Overview¶
- Paper Title: AdaST: Adaptive Coupling for Spatial-Temporal Forecasting
- Authors & Affiliations: Zhenyu Lei (University of Virginia), Chenghao Liu (Datadog), Yushun Dong (Florida State University), Qi R. Wang (Northeastern University), Jundong Li (University of Virginia)
- Paper Link: arXiv:2609.36119
- Code Repository: GitHub - LzyFischer/AdaST
- Core Keywords: Spatial-Temporal Forecasting, Coupling Structure Mismatch, Decompose-Recompose Paradigm, Heterogeneity-Aware Experts, Lightweight Spatial Mixer, Correlation-Informed Gating
Background and Motivation¶
Spatial-temporal forecasting serves as the technological cornerstone for managing critical real-world infrastructure systems, including intelligent traffic dispatch, climate and air quality monitoring, and electrical power grid load management. Real-world spatial-temporal data intrinsically exhibits two fundamental, intertwined dependencies: spatial correlation, which captures physical topology interactions among diverse locations or sensors, and temporal correlation, which characterizes the historical auto-evolutionary trajectory of each individual node. To model these interactions, modern research has developed a wide array of architectures, ranging from graph neural networks coupled with temporal convolutions (e.g., DCRNN, STGCN, Graph WaveNet) to spatial-temporal multi-head attention models (e.g., GMAN, STAEformer) and spatial-temporal diffusion frameworks. Despite their structural differences, these frameworks universally share a common inductive bias: they jointly couple spatial and temporal modeling uniformly at every layer, implicitly assuming that the underlying system maintains homogeneous, strong spatiotemporal coupling across all locations and time steps.
However, real-world physical systems exhibit substantially disparate coupling structures across diverse domains, categorizable into three distinct coupling regimes: 1. Weakly-Coupled Temporal-Dominated: Future trajectories are almost exclusively dictated by the node's own historical patterns, with negligible cross-location influence. For example, household power consumption is predominantly determined by internal domestic habits rather than neighboring activities; similarly, air quality readings are heavily governed by localized stationary emissions and micro-meteorology rather than immediate diffusion across distant stations. 2. Weakly-Coupled Spatial-Dominated: Adjacent neighboring conditions within immediate short-time windows dominate local dynamics, while long-term historical trends become largely secondary. For example, sudden highway accidents rapidly cause cascading upstream bottlenecks within minutes, primarily dictated by immediate neighboring flow spills rather than regular historical curves. 3. Strongly-Coupled: Historical temporal patterns and spatial network diffusion are tightly entangled. For example, during urban traffic peak hours, recurring commuter schedules interact dynamically with spatial congestion propagation across intersecting arterial corridors.
Through systematic empirical validation on controlled synthetic benchmarks, the paper demonstrates that a structural mismatch between rigid model coupling assumptions and intrinsic data patterns severely degrades predictive performance: forcing spatial coupling onto temporal-dominated data causes models to fit spurious cross-location dependencies; enforcing deep temporal modules on spatial-dominated scenarios incorporates misleading historical trends; even in strongly-coupled regimes, static architectures cannot accommodate fine-grained spatiotemporal shifts across different functional zones and peak/off-peak intervals.
Developing an adaptive spatial-temporal forecasting framework capable of dynamically modulating spatial and temporal modeling introduces three critical hurdles: (1) unknown coupling structure a priori across arbitrary real-world scenarios; (2) heterogeneous coupling dynamics exhibiting significant temporal periodicity and spatial functional variations within the same dataset; and (3) suboptimal spatial modeling operators, where static graphs lack dynamic adaptability, self-learned graphs overfit under noise, and full pairwise spatial attention incurs prohibitive quadratic complexity while amplifying spurious long-range links.
Core idea: AdaST establishes an adaptive decompose-recompose forecasting paradigm: it leverages multi-granularity heterogeneity-aware experts (incorporating spatial, daily, and weekly priors) to factorize input signals into temporal-specific, spatial-specific, and spatiotemporal-coupling components, processes each via role-aligned specialized modules, and adaptively recombines them using physical cosine correlation-informed gating to eliminate spurious cross-branch correlations while dynamically matching arbitrary coupling regimes.
Methodology¶
Overall Framework¶
AdaST executes a clean decompose-model-recompose pipeline, as illustrated in the following workflow diagram:
graph TD
%% Styling
classDef input fill:#e1f5fe,stroke:#0288d1,stroke-width:2px;
classDef decomp fill:#fff3e0,stroke:#f57c00,stroke-width:2px;
classDef model fill:#e8f5e9,stroke:#388e3c,stroke-width:2px;
classDef gate fill:#fce4ec,stroke:#c2185b,stroke-width:2px;
classDef output fill:#ede7f6,stroke:#512da8,stroke-width:2px;
X["Input ST Tensor X ∈ R^{N×T×D}"]:::input --> Emb["Linear Projection H0 ∈ R^{T×N×DH}"]:::input
subgraph Experts ["1. Heterogeneity-Aware Decoupling"]
Emb --> En["Spatial Expert (Node Embeddings En)"]:::decomp
Emb --> Eh["Time-of-Day Expert (Time Embeddings Eh)"]:::decomp
Emb --> Ew["Day-of-Week Expert (Day Embeddings Ew)"]:::decomp
Emb --> Ea["ST Joint Expert (ST Embeddings Ea)"]:::decomp
En & Eh & Ew & Ea --> Concat["Component Concat & Projection"]:::decomp
Concat --> Zt["Temporal Component H^{(t)}"]:::decomp
Concat --> Zs["Spatial Component H^{(s)}"]:::decomp
Concat --> Zst["ST Coupling Component H^{(st)}"]:::decomp
end
subgraph BranchModeling ["2. Role-Aligned Modeling"]
Zt --> TempModule["L-layer Temporal Attention (TempAttn)"]:::model
Zs --> SpatModule["L-layer Spatial Mixer (SpatMix)"]:::model
Zst --> STModule["L-layer Interleaved (TempAttn + SpatMix)"]:::model
TempModule --> Ht["Refined Temporal H_hat^{(t)}"]:::model
SpatModule --> Hs["Refined Spatial H_hat^{(s)}"]:::model
STModule --> Hst["Refined ST H_hat^{(st)}"]:::model
end
subgraph Recompose ["3. Correlation-Informed Adaptive Recomposition"]
Ht --> Ct["Temporal Cosine Sim C_t"]:::gate
Hs --> Cs["Spatial Cosine Sim C_s"]:::gate
Hst --> Cst["Joint Cosine Sim C_st"]:::gate
Ht & Hs & Hst --> Gating["Lightweight Gate g^{(·)} = σ(W · H_hat)"]:::gate
Gating & Ct & Cs & Cst --> Modulate["Modulated Gate g_tilde = g * [C]^α"]:::gate
Modulate --> Softmax["Softmax Normalized Weights (w^{(t)}, w^{(s)}, w^{(st)})"]:::gate
Softmax & Ht & Hs & Hst --> WeightedSum["Adaptive Sum H_hat = ∑ w^{(·)} ⊙ H_hat^{(·)}"]:::gate
end
WeightedSum --> Head["Temporal Projection & Head"]:::output
Head --> Pred["Future Prediction Y_hat ∈ R^{N×T'×D}"]:::output
The model ingests historical observations \(\mathbf{X} \in \mathbb{R}^{N \times T \times D}\), projecting them to an embedding space \(\mathbf{H}_0 \in \mathbb{R}^{T \times N \times D_0}\). Four heterogeneity experts decompose \(\mathbf{H}_0\) into three contextualized components: temporal-specific \(\mathbf{H}^{(t)}\), spatial-specific \(\mathbf{H}^{(s)}\), and spatiotemporally coupled \(\mathbf{H}^{(st)}\). Each component is refined independently by its dedicated modeling module. Finally, internal cosine similarity measurements assess feature signal quality to modulate a learned gating network, yielding dynamic mixture weights that combine the representations before final projection to \(\hat{\mathbf{Y}} \in \mathbb{R}^{N \times T' \times D}\).
Key Designs¶
AdaST breaks away from rigid joint-coupling conventions through three tightly integrated designs:
1. Heterogeneity-Aware Expert Decomposition: Coupling structures display substantial spatial and temporal variability across sensor positions and daily rhythms. Uniform decomposition cannot adapt to such multi-scale nuances. AdaST introduces four specialized expert types that inject learnable contextual embeddings: - Spatial Expert: Maintains node-specific embeddings \(\mathbf{E}_n \in \mathbb{R}^{N \times D_n}\) capturing location characteristics, concatenated with the projected input: \(\mathbf{H}_n = [\mathbf{H}_0 \,\|\, \tilde{\mathbf{E}}_n]\). - Time-of-Day Expert: Maintains diurnal step embeddings \(\mathbf{E}_h \in \mathbb{R}^{T_{\text{day}} \times D_h}\), indexing embeddings by daily timestamp \(t\): \(\mathbf{H}_h = [\mathbf{H}_0 \,\|\, \tilde{\mathbf{E}}_h[\mathbf{t}]]\). - Day-of-Week Expert: Maintains weekly embeddings \(\mathbf{E}_w \in \mathbb{R}^{T_{\text{week}} \times D_w}\), indexing embeddings by weekday \(d\): \(\mathbf{H}_w = [\mathbf{H}_0 \,\|\, \tilde{\mathbf{E}}_w[\mathbf{d}]]\). - Spatiotemporal Expert: Maintains full joint embeddings \(\mathbf{E}_a \in \mathbb{R}^{T \times N \times D_a}\) to capture dynamic unreduced interactions: \(\mathbf{H}_a = [\mathbf{H}_0 \,\|\, \mathbf{E}_a]\).
Each expert applies a dedicated projection head \(f_k: \mathbb{R}^{D_0 + D_k} \rightarrow \mathbb{R}^{\frac{3}{4}D_H}\) to produce three feature slices: $\([\mathbf{Z}_k^{(st)}; \mathbf{Z}_k^{(t)}; \mathbf{Z}_k^{(s)}] = f_k(\mathbf{H}_k), \quad \mathbf{Z}_k^{(\cdot)} \in \mathbb{R}^{T \times N \times \frac{1}{4}D_H}\)$ Outputs from all four experts are concatenated along the channel dimension to yield the final decomposed components: $\(\mathbf{H}^{(\cdot)} = [\mathbf{Z}_n^{(\cdot)} \,\|\, \mathbf{Z}_h^{(\cdot)} \,\|\, \mathbf{Z}_w^{(\cdot)} \,\|\, \mathbf{Z}_a^{(\cdot)}] \in \mathbb{R}^{T \times N \times D_H}, \quad (\cdot) \in \{st, t, s\}\)$
2. Role-Aligned Modeling with Lightweight Spatial Mixer: To prevent cross-component interference during representation learning, AdaST routes each component through specialized operators: - Temporal-Specific Branch: Models intra-series dynamics on \(\mathbf{H}^{(t)}\) using standard Multi-Head Self-Attention (MHA) along the temporal dimension, combined with LayerNorm and Feed-Forward Networks (FFN): $\(\mathbf{H}' = \text{MHA}(\mathbf{H}\mathbf{W}^Q, \mathbf{H}\mathbf{W}^K, \mathbf{H}\mathbf{W}^V), \quad \text{TempAttn}(\mathbf{H}) = \text{FFN}(\text{LN}(\mathbf{H} + \text{Dropout}(\mathbf{H}')))\)$ - Spatial-Specific Branch: Rather than relying on rigid predefined graphs or computationally heavy pairwise attention, AdaST designs a lightweight spatial mixer for \(\mathbf{H}^{(s)}\). It employs a global learnable transition matrix \(\mathbf{A} \in \mathbb{R}^{N \times N}\): $\(\text{SpatMix}(\mathbf{H}) = \text{FFN}\left(\mathbf{H} + \big(\text{softmax}(\mathbf{A}) \mathbf{H}^\top\big)^\top\right)\)$ The spatial mixer requires computing \(\text{softmax}(\mathbf{A})\) only once per forward pass, capturing all-pair spatial interactions without sample-level query-key scoring overhead, delivering a \(1.5\times\) to \(2.0\times\) speedup while preserving stability against spurious long-range links. - Spatiotemporal Coupling Branch: Stacks \(L\) layers that alternate \(\text{TempAttn}\) and \(\text{SpatMix}\) to capture complex spatial-temporal diffusion. This yields refined representations \(\hat{\mathbf{H}}^{(t)}, \hat{\mathbf{H}}^{(s)}, \hat{\mathbf{H}}^{(st)} \in \mathbb{R}^{T \times N \times D_H}\).
3. Correlation-Informed Adaptive Recomposition: To dynamically integrate the three components according to their reliability, AdaST combines a learned gating network with physical correlation strength measurements: First, a lightweight linear gate computes initial data-dependent scores: $\(g^{(\cdot)} = \sigma(W_{(\cdot)} \cdot \hat{\mathbf{H}}^{(\cdot)}) \in \mathbb{R}^{T \times N}, \quad (\cdot) \in \{t, s, st\}\)$ Next, physical continuity and spatial coherence are quantified via normalized cosine similarities across adjacent temporal slices and spatial nodes: $\(\mathcal{C}_t(\mathbf{H}^{(t)}) = \frac{1}{T-1}\sum_{i=1}^{T-1}\text{CosineSim}\left(\hat{\mathbf{H}}^{(t)}_{:, i, :, :}, \hat{\mathbf{H}}^{(t)}_{:, i+1, :, :}\right)\)$ $\(\mathcal{C}_s(\mathbf{H}^{(s)}) = \frac{1}{N-1}\sum_{j=1}^{N-1}\text{CosineSim}\left(\hat{\mathbf{H}}^{(s)}_{:, :, j, :}, \hat{\mathbf{H}}^{(s)}_{:, :, j+1, :}\right)\)$ $\(\mathcal{C}_{st}(\mathbf{H}^{(st)}) = \frac{1}{2}\big(\mathcal{C}_s(\mathbf{H}^{(st)}) + \mathcal{C}_t(\mathbf{H}^{(st)})\big)\)$ When a component is corrupted by spurious noise, its internal correlation metric drops. The gate scores are modulated via power scaling \(\alpha\): $\(\tilde{g}^{(\cdot)} = g^{(\cdot)} \cdot \left[\mathcal{C}_{(\cdot)}(\mathbf{H}^{(\cdot)})\right]^\alpha\)$ Applying Softmax yields normalized mixture weights \([w^{(t)}, w^{(s)}, w^{(st)}] = \text{softmax}(\tilde{g}^{(t)}, \tilde{g}^{(s)}, \tilde{g}^{(st)})\), and the recombined representation is formed as: $\(\hat{\mathbf{H}} = w^{(t)} \odot \hat{\mathbf{H}}^{(t)} + w^{(s)} \odot \hat{\mathbf{H}}^{(s)} + w^{(st)} \odot \hat{\mathbf{H}}^{(st)}\)$
Loss Function and Training Strategy¶
- Objective Function: The entire network is trained end-to-end to minimize the Mean Absolute Error (MAE / L1 loss) over all nodes across the \(T' = 12\) forecasting steps: $\(\mathcal{L} = \frac{1}{N \cdot T' \cdot D} \sum_{i=1}^N \sum_{t=1}^{T'} |\mathbf{Y}_{i,t} - \hat{\mathbf{Y}}_{i,t}|\)$
- Training Details & Hyperparameters:
- Hidden feature dimension \(D_H = 256\); expert embedding dimensions \(D_n = D_h = D_w = D_a = 24\);
- Network depth \(L = 3\) layers; multi-head temporal attention uses 4 heads;
- Correlation modulation weight \(\alpha = 0.1\) (selected via grid search over \(\{0.01, 0.05, 0.1, 0.5, 1.0\}\) on the validation set);
- Input horizon \(T = 12\) and prediction horizon \(T' = 12\);
- Optimized via Adam with an initial learning rate of 0.001, exponential decay, and batch size 16;
- Datasets are split into 60% training, 20% validation, and 20% testing using standard Z-score normalization.
Experimental Results and Analysis¶
Main Benchmark Comparisons¶
AdaST is extensively evaluated against 16 baselines across four spatial-temporal benchmarks: PurpleAir (temporal-dominated PM2.5 monitoring) and PEMS04, PEMS07, PEMS08 (spatially coupled highway traffic networks). Baselines encompass temporal-only methods (HI, DeepAR, NBeats), classical spatial-temporal GNNs (GWNet, DCRNN, AGCRN, STGCN, GTS, MTGNN, StemGNN), and modern decoupled/attention models (STNorm, D2STGNN, STID, STDN, HimNet, STAEformer).
Main Results Table (PurpleAir and PEMS04/07/08):
| Model Type | Baseline | PurpleAir (MAE / RMSE / MAPE) | PEMS04 (MAE / RMSE / MAPE) | PEMS07 (MAE / RMSE / MAPE) | PEMS08 (MAE / RMSE / MAPE) |
|---|---|---|---|---|---|
| Temporal-Only | HI | 3.430 / 5.983 / 72.52% | 42.35 / 61.66 / 29.92% | 49.03 / 71.18 / 22.75% | 36.66 / 50.45 / 21.63% |
| Temporal-Only | DeepAR | 0.994 / 1.817 / 32.48% | 20.64 / 32.35 / 14.28% | 22.00 / 35.44 / 9.31% | 16.80 / 26.38 / 10.66% |
| Temporal-Only | NBeats | 0.511 / 1.100 / 23.30% | 25.30 / 39.65 / 17.66% | 26.14 / 42.72 / 11.37% | 18.90 / 31.39 / 12.11% |
| Classical ST GNN | GWNet | 0.514 / 1.012 / 23.51% | 18.80 / 30.14 / 13.19% | 20.47 / 33.47 / 8.61% | 14.67 / 23.55 / 9.46% |
| Classical ST GNN | DCRNN | 0.656 / 1.268 / 26.99% | 19.63 / 31.26 / 13.59% | 21.16 / 34.14 / 9.02% | 15.22 / 24.17 / 10.21% |
| Classical ST GNN | AGCRN | 0.774 / 1.595 / 27.90% | 19.38 / 31.25 / 13.40% | 20.57 / 34.40 / 8.74% | 15.32 / 24.41 / 10.03% |
| Classical ST GNN | STGCN | 0.619 / 1.152 / 27.29% | 19.57 / 31.38 / 13.44% | 21.74 / 35.27 / 9.24% | 16.08 / 25.39 / 10.60% |
| Classical ST GNN | GTS | 0.617 / 1.157 / 29.90% | 20.96 / 32.95 / 14.66% | 22.15 / 35.10 / 9.38% | 16.49 / 26.08 / 10.54% |
| Classical ST GNN | MTGNN | 0.540 / 1.089 / 22.65% | 19.17 / 31.70 / 13.37% | 20.89 / 34.06 / 9.00% | 15.18 / 24.24 / 10.20% |
| Classical ST GNN | StemGNN | 0.512 / 1.038 / 22.89% | 22.98 / 36.00 / 16.56% | 22.50 / 36.41 / 9.57% | 16.90 / 26.30 / 11.89% |
| Modern Attention/Decoupled | STNorm | 0.574 / 1.102 / 23.69% | 18.96 / 30.98 / 12.69% | 20.50 / 34.66 / 8.75% | 15.41 / 24.77 / 9.76% |
| Modern Attention/Decoupled | D2STGNN | 0.618 / 1.495 / 22.71% | 18.32 / 29.89 / 12.51% | 19.49 / 32.59 / 8.09% | 14.10 / 23.36 / 9.33% |
| Modern Attention/Decoupled | STID | 0.563 / 1.137 / 26.69% | 18.42 / 29.91 / 12.39% | 19.61 / 32.69 / 8.31% | 14.21 / 23.35 / 9.32% |
| Modern Attention/Decoupled | STDN | 0.672 / 1.223 / 29.23% | 18.40 / 30.41 / 12.21% | 20.08 / 33.73 / 9.29% | 14.21 / 23.28 / 9.27% |
| Modern Attention/Decoupled | HimNet | 0.574 / 1.139 / 24.43% | 18.32 / 30.07 / 12.49% | 20.03 / 33.17 / 8.79% | 13.52 / 23.28 / 8.96% |
| Modern Attention/Decoupled | STAEformer | 0.563 / 1.060 / 22.98% | 18.35 / 30.73 / 12.25% | 19.46 / 33.56 / 8.41% | 13.56 / 23.72 / 8.92% |
| Adaptive Decompose-Recompose | AdaST (Ours) | 0.489 / 0.994 / 22.43% | 18.28 / 29.87 / 12.17% | 19.16 / 32.54 / 8.02% | 13.45 / 23.17 / 8.85% |
Key takeaways from the main results: 1. Consistent SOTA Accuracy: AdaST achieves the best results across all four benchmarks on every metric, delivering average improvements of 1.7% in MAE, 0.7% in RMSE, and 0.8% in MAPE over the strongest baselines. 2. Substantial Gain on Weakly-Coupled Data: The advantage is most noticeable on PurpleAir, where AdaST outperforms the best baseline by 4.3% in MAE. On this dataset, purely temporal baselines like NBeats comfortably outperform complex ST-GNNs because forced spatial coupling introduces spurious cross-sensor noise. AdaST effectively suppresses unnecessary spatial branches via low spatial gate scores, avoiding negative transfer.
Ablation Study and Interpretability¶
To analyze module necessity, systematic ablations were conducted on PurpleAir and PEMS07.
Ablation Study Table:
| Ablation Variant | PurpleAir MAE | PurpleAir RMSE | PurpleAir MAPE | PEMS07 MAE | PEMS07 RMSE | PEMS07 MAPE |
|---|---|---|---|---|---|---|
| w/o Spatial Expert (\(\mathbf{E}_n\)) | 0.519 | 1.072 | 23.79% | 19.44 | 32.84 | 8.49% |
| w/o Time-of-Day Expert (\(\mathbf{E}_h\)) | 0.495 | 0.999 | 22.50% | 20.39 | 34.14 | 8.68% |
| w/o Day-of-Week Expert (\(\mathbf{E}_w\)) | 0.498 | 1.023 | 22.85% | 19.21 | 32.63 | 8.07% |
| w/o ST Joint Expert (\(\mathbf{E}_a\)) | 0.516 | 1.075 | 23.47% | 20.16 | 33.23 | 11.26% |
| Replace with Spatial Attention (SpaAtt) | 0.518 | 1.072 | 23.42% | 19.35 | 32.93 | 8.11% |
| w/o Correlation Modulation (\(w/o \ \mathcal{C}\)) | 0.513 | 1.067 | 23.25% | 19.85 | 33.02 | 10.58% |
| w/o Adaptive Gating (\(w/o \ g\)) | 0.526 | 1.088 | 24.00% | 20.29 | 33.35 | 9.57% |
| Full AdaST Model | 0.489 | 0.994 | 22.43% | 19.16 | 32.54 | 8.02% |
Key insights from ablation and interpretability analysis: - Domain-Specific Expert Importance: On PurpleAir, removing the spatial expert \(\mathbf{E}_n\) causes the largest performance drop (MAE 0.489 \(\rightarrow\) 0.519), indicating that location-specific microclimate biases dominate sensor variation. Conversely, on PEMS07, omitting the time-of-day expert \(\mathbf{E}_h\) triggers the steepest drop (MAE 19.16 \(\rightarrow\) 20.39), highlighting the dominant role of diurnal commuting cycles in traffic. - Spatial Mixer Efficiency & Robustness: Replacing the lightweight spatial mixer with full spatial attention (SpaAtt) harms accuracy significantly on PurpleAir, proving that unconstrained attention overfits spurious long-range links in weakly-coupled regimes. Moreover, runtime profiling shows the spatial mixer is \(1.5\times\) to \(2.0\times\) faster per epoch than spatial attention. - Necessity of Correlation-Informed Modulation: Eliminating gating entirely (\(w/o \ g\)) leads to the worst degradation across both datasets. Retaining gating without correlation modulation (\(w/o \ \mathcal{C}\)) also degrades performance significantly, confirming that explicit cosine-similarity regularization prevents the gating network from allocating weight to noisy components.
Limitations and Insights¶
Limitations¶
- Parameter and Memory Overhead on Large Graphs: While each sub-module is lightweight, maintaining four expert branches and three parallel modeling streams incurs higher activation memory during training compared to simple single-stream models, which may limit scalability on massive graphs with tens of thousands of nodes.
- Short-Term Horizon Focus: Experiments are currently conducted within standard short-term forecasting horizons (\(T = T' = 12\) steps). Assessing whether decomposed representations stay stable over long-term prediction spans (e.g., hundreds of steps) remains an open area for future exploration.
Key Takeaways and Future Directions¶
- Rethinking Uniform Coupling: Rather than blindly engineering increasingly dense spatial-temporal attention or differential equations that force joint coupling everywhere, AdaST demonstrates that disentangling orthogonal dynamics and adaptively recoupling them according to data-driven signals yields superior generalization and robustness.
- Broader Applicability to Multivariate Time Series: As demonstrated on non-spatial benchmarks (ETTh1, ExchangeRate), treating feature channels as pseudo-nodes allows AdaST to generalize beyond physical sensor grids, presenting a promising architectural blueprint for unified time-series foundation models.