Pith. sign in

REVIEW 5 major objections 6 minor 36 references

Ada-TransGNN: An Air Quality Prediction Model Based On Adaptive Graph Convolutional Networks

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Learning the station graph from data plus a spatial-autocorrelation task lowers AQI forecast error at every horizon tested.

desk verdict Plausible architecture and a useful new dataset, but the SOTA claim rests on single-run comparisons and the auxiliary-task math is too broken to support the paper's main attribution. read the letter →

arxiv 2508.17867 v2 pith:GVXNNUEE submitted 2025-08-25 cs.LG cs.AI

Dan Wang , Feng Jiang , Zhanquan Wang This is my paper
classification cs.LGcs.AI
keywords airqualitypredictionadaptivegraphstructurelearningauxiliarytaskMorancoefficientspatio-temporalforecastingconvolutionalnetworksTransformerMete-airdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes Ada-TransGNN, a Transformer-based model that learns the graph connecting air quality monitoring stations from the data instead of assuming a fixed geographic network, and tests whether this plus an auxiliary spatial-autocorrelation task improves AQI forecasts. The graph is built from a macro component capturing stable relations and a micro component capturing short-term changes, then combined as a learned adjacency matrix. Spatio-temporal blocks of multi-head attention and Chebyshev graph convolution extract features, and an auxiliary head predicts a Moran-coefficient-based spatial signal whose loss is uncertainty-weighted against the main prediction loss. On a 209-city dataset and a new 120-station Mete-air dataset, the model reports the lowest MAE and RMSE across all tested horizons, including 72-hour forecasts. If right, the recipe is: learn the spatial graph adaptively, supervise it with spatial autocorrelation, and use attention plus graph convolution jointly.

What carries the argument

The load-bearing object is the learned adjacency matrix A^(1) = Norm(ReLU(A_ma · A_mi)), which replaces the fixed geographic graph: A_ma is a macro-structure matrix generated from node attributes and added to an initial distance-based A_0, while A_mi is a micro-structure matrix produced by applying two convolutions to the input node features. This matrix is what the Chebyshev graph convolution operates on. The other essential mechanism is the auxiliary Moran-coefficient task: a standard statistic of spatial autocorrelation is computed from node features and used as a second prediction target, so the network is explicitly trained to represent spatial structure, with the auxiliary loss weighte

What would settle it

Compute a standard Moran's I from the released code's auxiliary outputs on Dataset 1 and compare with Eqs. (10)–(11) as written after supplying the undefined terms; if they do not match, the ablation gain cannot be attributed to spatial autocorrelation. Alternatively, fix the formulas and rerun the no-Moran ablation: if the MAE gap shrinks to noise, the paper's central explanation is falsified.

Watch

Extended reading notes

Core claim

Ada-TransGNN is presented as a spatio-temporal forecasting architecture whose main departure from prior work is an adaptive graph structure learning module. Instead of a predefined adjacency matrix based on geography, the model computes A^(1) = Norm(ReLU(A_ma · A_mi)): a macro adjacency learned from node attributes, a micro adjacency learned from time-varying node features by two convolutional layers, and a normalized combination. Stacked spatio-temporal blocks—multi-head attention followed by Chebyshev graph convolution with residual connections—then extract features, and an auxiliary task predicts a Moran-coefficient target to inject spatial autocorrelation information. The two losses are

Load-bearing premise

The claim that the Moran-coefficient auxiliary task drives the improvement depends on the task being implemented as written, but the paper's Eqs. (10)–(11) contain undefined symbols (z_ij, l), a reused summation index, and a reference to Eq. (9) that is actually the Chebyshev convolution, so the credited mechanism is not fully specified.

Editorial extensions

If this is right

  • On Dataset 1 (209 cities, 6-hour forecasting), Ada-TransGNN has the lowest MAE and RMSE at every horizon from 1h to 6h; at 6h the improvement over the strongest baseline is about 0.3 MAE and 0.65 RMSE.
  • On Mete-air (120 stations, 72-hour forecasting), it reports MAE 11.60 and RMSE 16.04, lower than all baselines, with the next-best model at 12.34 MAE and 17.29 RMSE.
  • Ablation results indicate the Moran auxiliary task and the macro/micro graph-learning modules each contribute to the reported performance, with the Moran task identified as the largest single source of gain.
  • A single architecture handles both short-term and long-term forecasts by converting hidden dimensions directly, which the paper argues avoids error accumulation from recursive multi-step decoding.
  • The Mete-air dataset is released for future benchmarking alongside the code.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the auxiliary task is a generic spatial-autocorrelation signal, the same uncertainty-weighted auxiliary-loss recipe could transfer to other geospatial time-series problems, such as traffic or weather, where nearby stations are correlated.
  • The reported margins over the strongest baselines are small (roughly 0.08–0.3 MAE on Dataset 1); a reader should treat the superiority claim as provisional until variance across seeds and statistical significance are reported, which the paper does not do.
  • Given the inconsistency in the printed Moran-coefficient equations, the credited mechanism can only be evaluated from the released code; if the implemented formula is a standard Moran's I, the method is reproducible from the code rather than from the text.
  • The adaptive adjacency could make the model applicable where station locations are unreliable or missing, since the graph is inferred from pollutant and meteorological features rather than coordinates.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes Ada-TransGNN, a Transformer-GNN hybrid for multi-step air quality index (AQI) forecasting. The architecture consists of a data embedding layer, an adaptive graph structure learning module with macro and micro components, L spatio-temporal blocks combining multi-head temporal attention with Chebyshev graph convolution, and an auxiliary task that uses Moran coefficients to capture spatial autocorrelation. The auxiliary loss is combined with the main prediction loss through an uncertainty-based weighting scheme. The authors evaluate on a 209-city benchmark dataset and a newly introduced 120-station dataset (Mete-air), comparing against nine baselines and reporting lower MAE and RMSE at 1--6 hour horizons (Dataset 1) and at 72 hours (Dataset 2). They also provide ablation and parameter sensitivity experiments, and state that code and datasets are available on GitHub. The central claim is that Ada-TransGNN outperforms existing state-of-the-art models in both short-term and long-term air quality forecasting.

Significance. The paper addresses a practically important problem and introduces a new real-world dataset (Mete-air) along with a broad baseline comparison on held-out test sets. The adaptive graph learning and auxiliary spatial-autocorrelation task are plausible mechanisms for improving spatio-temporal forecasting, and the authors make code and data availability a stated part of the contribution, which is a strength. If the reported results are confirmed, the contribution would be a useful incremental improvement over existing air quality forecasting models. However, the current manuscript contains several load-bearing formal gaps: key equations for the Moran coefficient, the loss weighting, and the graph learning module are either undefined or internally inconsistent, and the headline empirical claim rests on single-run results without variance estimates or significance tests. The significance of the work is therefore conditional on substantial revision and re-evaluation.

major comments (5)
  1. [§4.4, Tables 2 and 3] The claim that Ada-TransGNN outperforms all baselines rests on a single run. On Dataset 1 (Table 2), the 1-hour MAE margin over GAGNN is 5.75 vs 5.83 (Δ=0.08, about 1.4%); on Dataset 2 (Table 3), the margin over FCSTGNN is 11.60 vs 12.34 (Δ=0.74, about 6%). No error bars, number of seeds, or significance tests are reported, and baseline hyperparameter tuning is not described. These gaps are small enough that they may fall within run-to-run variation. The authors should report mean±std over at least five independent runs and, for the headline comparisons, a paired test (e.g., Diebold-Mariano or paired bootstrap over test samples). Without this, the SOTA claim is not supported by the evidence.
  2. [§3.5, Eqs. (10)-(11)] The Moran coefficient formulas are not well-defined. Eq. (10) uses symbols z_ij and z_il with no definitions, and the index l is never introduced. Eq. (11) reuses the summation index j for both the outer and inner sums, and z_ij appears again without definition. The standard Moran's I formula is I = (N/S0) * (Σ_i Σ_j w_ij z_i z_j)/(Σ_i z_i^2), which is different from what appears here. Since the ablation in Fig. 3 attributes a major improvement to the Moran auxiliary task, the exact definition used in the implementation must be provided. Please correct the formulas, define every symbol, and confirm consistency with the released code.
  3. [§3.6, text after Eq. (12)] The text states: 'We calculate the spatial weight y_moran of all monitoring point eigenvalues x_i according to Eq. (9) in each training step.' Eq. (9) is the Chebyshev graph convolution formula, not a spatial weight or a Moran coefficient. This cross-reference makes it impossible to determine what the auxiliary task actually computes. The authors should replace the reference with the correct equation, define y_moran before it is used in Eq. (13), and specify how the Moran coefficients enter the auxiliary regression target.
  4. [§3.6, Eqs. (14)-(15)] The uncertainty-weighting formula is incoherent as written. Eq. (14) is a convex combination λ·MSE(main) + (1−λ)·MSE(aux), but Eq. (15) sets λ = 1/(2σ²). Standard uncertainty weighting for two tasks uses two separate uncertainties, with weights 1/(2σ₁²) and 1/(2σ₂²) and additional log σ terms; a single σ cannot produce a convex weight in this way, and λ > 1 would make 1−λ negative. Furthermore, §4.6 reports sweeping λ over {0.1, 0.3, 0.5, 0.7} and finding 0.5 optimal, which contradicts the statement that λ is learned automatically. Please clarify whether λ is a tuned hyperparameter or a learned parameter, and give the actual loss formula used.
  5. [§3.3, Eqs. (2)-(5)] The adaptive graph structure learning module is under-specified. The function f in Eq. (2) is never defined. Eq. (4) applies Conv2d to x_i ∈ R^{N×T×D} but does not state how the output becomes an N×N adjacency matrix. Eq. (5) multiplies A_ma and A_mi without specifying how their shapes are reconciled, and the 'subsumption layer' mentioned in the text is not described. The initial adjacency matrix A0 is also not defined in detail. Because this module is a central contribution and is ablated in Fig. 3, the authors should provide precise tensor shapes, operation sequences, and any normalization steps.
minor comments (6)
  1. [§3.4, Eq. (9)] The text calls the Chebyshev convolution 'domain-based,' but Chebyshev spectral graph convolution is a spectral method. This terminology should be corrected.
  2. [§2, References] There are duplicate and near-duplicate references for DCRNN (refs [6] and [7]) and for STGCN (refs [8] and [32]). The reference list should be cleaned and renumbered.
  3. [§4.3] The experimental setting lists the optimizer, batch size, learning rate, and number of attention heads, but omits random seeds, number of runs, validation-based early stopping criteria, and the hyperparameter tuning protocol for baselines. These details are necessary for reproducibility.
  4. [§4.6] In the parameter sensitivity discussion, the authors state that λ = 0.5 gives the best performance, but the list of tested values is incomplete (0.1, 0.3, 0.5, 0.7; no 0.9). Please report the full range and clarify the relationship to the automatic uncertainty weighting described in §3.6.
  5. [§3.4, Eqs. (6)-(8)] Eqs. (6) and (8) are identical. The redundancy is confusing; Eq. (8) should either be removed or clearly defined as the per-head computation with head-specific projections.
  6. [Figures 3 and 4] The captions do not explain the axis labels, units, or what is being compared. Please expand the captions and ensure the figures are legible in the printed version.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the SOTA claim is supported by held-out test evaluation against external baselines; no prediction reduces to fitted inputs by construction.

full rationale

Walking the derivation chain, the paper's central claim is empirical: Ada-TransGNN is trained on past AQI/meteorological data and evaluated on held-out test portions of Dataset 1 and Dataset 2 against published external baselines (ARIMA, LSTM, STGCN, DMSTGCN, FCSTGNN, GMAN, ASTGCN, STAGCN, GAGNN). There is no hidden equivalence whereby a fitted parameter is renamed as a prediction: the adaptive graph adjacency A^(1)=Norm(ReLU(A_ma·A_mi)) and the uncertainty weight λ are learned from training data, which is standard model fitting rather than circularity. The auxiliary Moran task uses input features to construct a spatial-autocorrelation target; even if the equations are internally inconsistent (Eq. 10 uses undefined symbols z_ij and z_il, Eq. 11 reuses the summation index j, and Section 3.6 says the spatial weight is computed according to Eq. 9, which is the Chebyshev convolution, not a Moran formula), this is a correctness/implementation ambiguity, not a case of the prediction being defined in terms of the target. Similarly, Section 3.6 states λ is learned automatically via Eq. 15, while Section 4.6 treats λ as a tuned hyperparameter with best value 0.5; this is a reporting inconsistency, not circularity. No load-bearing conclusion rests on a self-citation: the only dataset-1 citation [17] is to GAGNN, an external baseline, and no uniqueness theorem or ansatz is imported from the authors' own prior work to force the model choice. The absence of error bars and significance tests is a reproducibility concern about the strength of the SOTA claim, but it does not make the evaluation circular.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical entities, particles, or conserved quantities. Its free parameters are conventional machine-learning hyperparameters and a loss-balance weight. The axioms are standard math tools plus domain assumptions about air quality data and several architecture decisions unique to this paper. The main burden rests on the ad-hoc claims that the macro-micro graph learner and the Moran auxiliary task improve prediction.

free parameters (4)
  • Auxiliary task weight lambda = 0.5 in sensitivity sweep; claimed to be learned via sigma-squared in Eq. (15)
    Trade-off weight in the combined loss. The paper gives two incompatible accounts: a fixed scalar tuned in Fig. 4 and a learned uncertainty weight in Eq. (15).
  • Number of attention heads h = 2 (default); tuned to 1, 2, 5, 10
    Hyperparameter of the multi-head attention layer, tuned in parametric sensitivity experiments (Section 4.6).
  • Number of spatio-temporal blocks L = 3 (default); tuned to 1, 3, 5, 6
    Depth of the stacked ST-blocks, tuned in parametric sensitivity experiments (Section 4.6).
  • Chebyshev polynomial order K
    Appears in Eq. (9) as the truncation order of Chebyshev graph convolution; no value is reported anywhere in the paper.
assumptions (6)
  • standard math Chebyshev polynomial graph convolution provides an effective spatial operator on monitoring-station graphs (Eq. 9).
    Invoked from Defferrard et al. (ref [26]) without proof.
  • standard math Multi-head scaled dot-product attention captures temporal dependencies (Eqs. 6-8).
    Standard Transformer machinery from Vaswani et al. (ref [23]).
  • domain assumption Air quality observations across stations form a graph signal that contains the predictive information needed for forecasting.
    The problem formulation in Section 3.1 assumes the graph G and node features X_t fully describe the prediction task.
  • ad hoc to paper The macro-micro decomposition of graph learning (Eqs. 2-5) is a valid and useful architecture for discovering inter-station dependencies.
    No theoretical or empirical justification is given beyond the ablation study in Section 4.5.
  • ad hoc to paper Using Moran coefficients as an auxiliary regression target improves the main prediction task.
    The auxiliary task loss is introduced in Section 3.5 with no derivation showing why it should help the primary objective.
  • ad hoc to paper The uncertainty-weighting formula lambda = 1/(2*sigma-squared) yields a valid task combination in Eq. (14).
    The formula is stated without derivation and is inconsistent with the complementary weighting (1-lambda) used in Eq. (14).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ada-TransGNN: An Air Quality Prediction Model Based On Adaptive Graph Convolutional Networks." pith.science (2026). https://pith.science/paper/GVXNNUEE

@misc{pith2026250817867,
  author       = {Pith},
  title        = {Pith review of: Ada-TransGNN: An Air Quality Prediction Model Based On Adaptive Graph Convolutional Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GVXNNUEE}},
  note         = {Machine review of arXiv:2508.17867}
}
read the original abstract

Accurate air quality prediction is becoming increasingly important in the environmental field. To address issues such as low prediction accuracy and slow real-time updates in existing models, which lead to lagging prediction results, we propose a Transformer-based spatiotemporal data prediction method (Ada-TransGNN) that integrates global spatial semantics and temporal behavior. The model constructs an efficient and collaborative spatiotemporal block set comprising a multi-head attention mechanism and a graph convolutional network to extract dynamically changing spatiotemporal dependency features from complex air quality monitoring data. Considering the interaction relationships between different monitoring points, we propose an adaptive graph structure learning module, which combines spatiotemporal dependency features in a data-driven manner to learn the optimal graph structure, thereby more accurately capturing the spatial relationships between monitoring points. Additionally, we design an auxiliary task learning module that enhances the decoding capability of temporal relationships by integrating spatial context information into the optimal graph structure representation, effectively improving the accuracy of prediction results. We conducted comprehensive evaluations on a benchmark dataset and a novel dataset (Mete-air). The results demonstrate that our model outperforms existing state-of-the-art prediction models in short-term and long-term predictions.

Figures

Figures reproduced from arXiv: 2508.17867 by the authors.

Figure 1
Figure 1. spatio - temporal dependences To achieve more efficient air quality prediction, existing studies have made significant progress based on graph neural networks and Transformer. How￾ever, these approaches underutilize the complex spatio-temporal correlations. For example, approaches[6-8] based STGNN typically learn spatial relationships through GNNs and capture temporal dependencies in combination with RNNs or CNNs. H… view at source ↗
Figure 2
Figure 2. illustrates the framework of the Ada-TransGNN model, which mainly consists of a data embedding layer, an adaptive graph structure learning module, L spatio-temporal blocks, and two output layers. We will carefully describe each module in the following [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Ablation research 4.6 Parametric Sensitivity experiment We conducted an experiment to analyze the impact of three key hyperparame￾ters: the number of heads of the multi-head attention mechanism h, the number of spatio-temporal blocks b, and the weights of the two output layers in the loss function λ. In [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Parametric Sensitivity research 5 Conclusion In this paper, an air quality index prediction model for Ada-TransGNN is pro￾posed. To effectively model the spatio-temporal dependence of complex air qual￾ity data, the framework is modeled by stacking spatio-temporal block…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 7 linked inside Pith

  1. [1]

    Associations between long-term ozone exposure and small airways function in Chinese young adults: a longitudinal cohort study [J]

    Feng S., Yang L., Dou S.,et al. Associations between long-term ozone exposure and small airways function in Chinese young adults: a longitudinal cohort study [J]. Respiratory Research, 2024, 25(1): 105

    2024

  2. [2]

    Avoidable mortality due to long-term exposure to PM2

    Rodriguez-Villamizar L A, Belalcazar-Ceron L C, Castillo M P, et al. Avoidable mortality due to long-term exposure to PM2. 5 in Colombia 2014–2019[J]. Envi- ronmental Health, 2022, 21(1): 137

    2014

  3. [3]

    Wang, S., Hao, J.: Air quality management in China: Issues, challenges, and op- tions. J. Environ. Sci., 24(1), 2 - 13 (2012). 14 D. Wang et al

    2012

  4. [4]

    Zhai, S., Jacob, D.J., Wang, X., et al.: Fine particulate matter (PM2.5) trends in China, 2013 - 2018: Separating contributions from anthropogenic emissions and meteorology. Atmos. Chem. Phys., 19(16), 11031 - 11041 (2019)

    2013

  5. [5]

    Guo, B., Wang, X., Pei, L., et al.: Identifying the spatiotemporal dynamic of PM2.5 concentrations at multiple scales using geographically and temporally weighted regression model across China during 2015 - 2018. Sci. Total Environ., 751, 141765 (2021)

    2015

  6. [6]

    arXiv preprint arXiv:1707.01926 (2017)

    Li, Y., Yu, R., Shahabi, C., et al.: Diffusion convolutional recurrent neural network: Data - driven traffic forecasting. arXiv preprint arXiv:1707.01926 (2017)

    arXiv 2017

  7. [7]

    Li, Y., Yu, R., Shahabi, C., Liu, Y.: Diffusion convolutional recurrent neural net- work: data - driven traffic forecasting.16 (2018)

    2018

  8. [9]

    arXiv preprint arXiv:2001.02908 (2020)

    Xu, M., Dai, W., Liu, C., et al.: Spatial - temporal transformer networks for traffic flow forecasting. arXiv preprint arXiv:2001.02908 (2020)

    arXiv 2001

Show all 36 references
  1. [10]

    In: The Eleventh International Con- ference on Learning Representations, 2023

    Zhang, Y., Yan, J.: Crossformer: Transformer utilizing cross - dimension depen- dency for multivariate time series forecasting. In: The Eleventh International Con- ference on Learning Representations, 2023

    2023

  2. [11]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(12): 11106–11115

    Zhou, H., Zhang, S., Peng, J., et al.: Informer: Beyond efficient transformer for long sequence time - series forecasting. In: Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(12): 11106–11115

    2021

  3. [12]

    Appel, K.W., Bash, J.O., Fahey, K.M., et al.: The Community Multiscale Air Quality (CMAQ) model versions 5.3 and 5.3.1: system updates and evaluation. Geosci. Model Dev., 14(5), 2867 - 2897 (2021)

    2021

  4. [13]

    Vautard, R., Moran, M.D., Solazzo, E., et al.: Evaluation of the meteorological forc- ing used for the Air Quality Model Evaluation International Initiative (AQMEII) air quality simulations. Atmos. Environ., 53, 15 - 37 (2012)

    2012

  5. [14]

    Expert Syst

    Zhang, B., Zou, G., Qin, D., et al.: RCL - Learning: ResNet and convolutional long short - term memory - based spatiotemporal air pollutant concentration prediction model. Expert Syst. Appl., 207, 118017 (2022)

    2022

  6. [15]

    arXiv preprint arXiv:2101.04264 (2021)

    Xu, J., Chen, L., Lv, M., et al.: HighAir: A hierarchical graph neural network - based air quality forecasting method. arXiv preprint arXiv:2101.04264 (2021)

    2021

  7. [16]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, 2023, 37(12): 14329 - 14337

    Liang, Y., Xia, Y., Ke, S., et al.: Airformer: Predicting nationwide air quality in China with transformers. In: Proceedings of the AAAI Conference on Artificial Intelligence, 2023, 37(12): 14329 - 14337

    2023

  8. [17]

    ACM Trans

    Chen, L., Xu, J., Wu, B., et al.: Group - aware graph neural network for nationwide city air quality forecasting. ACM Trans. Knowl. Discov. Data, 18(3), 1 - 20 (2023)

    2023

  9. [18]

    In: Proceedings of the AAAI conference on artificial intelligence, 2017, 31(1)

    Zhang, J., Zheng, Y., Qi, D.: Deep spatio - temporal residual networks for city- wide crowd flows prediction. In: Proceedings of the AAAI conference on artificial intelligence, 2017, 31(1)

    2017

  10. [19]

    arXiv preprint arXiv:1810.05749 (2018)

    Zhang, C., Ren, M., Urtasun, R.: Graph hypernetworks for neural architecture search. arXiv preprint arXiv:1810.05749 (2018)

    2018 arXiv

  11. [20]

    arXiv preprint arXiv:1906.00121 (2019)

    Wu, Z., Pan, S., Long, G., et al.: Graph wavenet for deep spatial - temporal graph modeling. arXiv preprint arXiv:1906.00121 (2019)

    1906 arXiv

  12. [21]

    In: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 2020, pp

    Wu, Z., Pan, S., Long, G., et al.: Connecting the dots: Multivariate time series forecasting with graph neural networks. In: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 2020, pp. 753 - 763. Ada-TransGNN for Air Quality Pred...

    2020

  13. [22]

    In: Proceedings of the web conference 2020, 2020, pp

    Wang, X., Ma, Y., Wang, Y., et al.: Traffic flow prediction via spatial temporal graph neural network. In: Proceedings of the web conference 2020, 2020, pp. 1082 - 1092

    2020

  14. [23]

    Vaswani, A., Shazeer, N., Parmar, N., et al.: Attention is all you need. Adv. Neural Inf. Process. Syst., 30 (2017)

    2017

  15. [24]

    arXiv preprint arXiv:1312.6203 (2013)

    Bruna, J., Zaremba, W., Szlam, A., et al.: Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203 (2013)

    2013 arXiv

  16. [25]

    arXiv preprint arXiv:1609.02907 (2016)

    Kipf, T.N., Welling, M.: Semi - supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  17. [26]

    In: Proceedings of the 30th Interna- tional Conference on Neural Information Processing Systems (NIPS’16), 2016, pp

    Defferrard, M., Bresson, X., Vandergheynst, P.: Convolutional neural networks on graphs with fast localized spectral filtering. In: Proceedings of the 30th Interna- tional Conference on Neural Information Processing Systems (NIPS’16), 2016, pp. 3844 - 3852

    2016

  18. [27]

    Biometrika, 37(1/2), 17 - 23 (1950)

    Moran, P.A.P.: Notes on continuous stochastic phenomena. Biometrika, 37(1/2), 17 - 23 (1950)

    1950

  19. [28]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp

    Kendall, A., Gal, Y., Cipolla, R.: Multi - task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7482 - 7491

    2018

  20. [29]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019, pp

    Shen, T., Zhou, L., Luo, Z., et al.: Self - supervised learning of depth and motion under photometric inconsistency. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, 2019, pp. 0 - 0

    2019

  21. [30]

    In: IJCAI, 2018, pp

    Liang, Y., Ke, S., Zhang, J., et al.: Geoman: Multi - level attention networks for geo - sensory time series prediction. In: IJCAI, 2018, pp. 3428 - 3434

    2018

  22. [31]

    In: Supervised sequence labelling with recurrent neural networks, 2012, pp

    Graves, A.: Long short - term memory. In: Supervised sequence labelling with recurrent neural networks, 2012, pp. 37 - 45

    2012

  23. [32]

    arXiv preprint arXiv:1709.04875 (2017)

    Yu, B., Yin, H., Zhu, Z.: Spatio - temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875 (2017)

    2017 arXiv

  24. [33]

    In: Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp

    Han, L., Du, B., Sun, L., et al.: Dynamic and multi - faceted spatio - temporal deep learning for traffic speed forecasting. In: Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp. 547 - 555

    2021

  25. [34]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(14): 15715 - 15724

    Wang, Y., Xu, Y., Yang, J., et al.: Fully - Connected Spatial - Temporal Graph for Multivariate Time - Series Data. In: Proceedings of the AAAI Conference on Artificial Intelligence, 2024, 38(14): 15715 - 15724

    2024

  26. [35]

    In: Proceedings of the AAAI conference on artificial intelligence, 2020, 34(01): 1234 - 1241

    Zheng, C., Fan, X., Wang, C., et al.: Gman: A graph multi - attention network for traffic prediction. In: Proceedings of the AAAI conference on artificial intelligence, 2020, 34(01): 1234 - 1241

    2020

  27. [36]

    In: Proceedings of the AAAI conference on artificial intelligence, 2019, 33(01): 922 - 929

    Guo, S., Lin, Y., Feng, N., et al.: Attention based spatial - temporal graph convolu- tional networks for traffic flow forecasting. In: Proceedings of the AAAI conference on artificial intelligence, 2019, 33(01): 922 - 929

    2019

  28. [37]

    Mathematics, 10(9), 1599 (2022)

    Gu, Y., Deng, L.: STAGCN: Spatial–temporal attention graph convolution network for traffic forecasting. Mathematics, 10(9), 1599 (2022)

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.