Pith. sign in

REVIEW 5 major objections 6 minor 30 references

An Interpretable Implicit-Based Approach for Modeling Local Spatial Effects: A Case Study of Global Gross Primary Productivity

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

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

Pith's one-line read A dual-branch encoder-decoder that jointly learns location-invariant features and location-specific spatiotemporal conditions predicts global gross primary productivity at RMSE 0.836, beating tabular and GWR-family baselines while…

desk verdict A coherent architecture and a large new benchmark, but the headline RMSE gain may just be emulating the PML V2 model product; independent validation is needed before the predictive claim is taken seriously. read the letter →

arxiv 2502.06170 v1 pith:IHOEFUFW submitted 2025-02-10 cs.CV cs.AIcs.LG

Siqi Du , Hongsheng Huang , Kaixin Shen , Ziqi Liu , Shengjun Tang This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords spatiotemporalheterogeneitygrossprimaryproductivitydual-branchencoder-decodergraphconvolutionalnetworkLSTMtransformerinterpretabilityClimate2GPP
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

This paper argues that geographic machine learning problems, where the relationship between variables changes across space and time, are best solved by decomposing the task into two learned components: a location-invariant representation of the underlying physical law and a location-specific spatiotemporal condition. The authors build a dual-branch encoder-decoder network that learns these two components jointly, using graph convolution and LSTM over a spatiotemporal conditional graph for the local branch and linear self-attention for the global branch. Applied to 50 million global samples of climate and land cover data, the model predicts vegetation gross primary productivity for 2020 with RMSE 0.836, outperforming LightGBM, TabNet, and GWR-family baselines. It also produces per-location, per-time importance weights that visualize how the dominant drivers of GPP vary across biomes. A sympathetic reader would take the central claim to be that explicit joint modeling of global patterns and local differences, rather than local coefficient fitting, is what makes the method accurate and interpretable.

What carries the argument

The load-bearing object is the Spatiotemporal Conditional Graph (STCG), a graph whose nodes are cluster centers of the global land grid at each time step, with node2vec embeddings plus rotary position embeddings, edge weights from a log-Gaussian kernel on 3D spherical coordinates, and KNN adjacency on the sphere to make the graph cyclic. Graph convolution aggregates the hidden condition vectors spatially, 1D convolution aggregates them temporally, and an LSTM captures long-term dependencies, producing the implicit condition vector that conditions the decoder. The second branch is a dual self-attention tabular encoder that extracts location-invariant features from the meteorological time series, and the decoder uses cross-attention with the condition vector as value to predict GPP and feature importance weights. This combination is what allows the architecture to separate what is generally true from what is locally different, and to express the latter as a continuous, time-varying condition rather than a set of local regression coefficients.

What would settle it

Compare the model's GPP predictions against independent field-based carbon uptake measurements from eddy-covariance tower sites that were not used in training, and see whether the RMSE advantage over tabular baselines persists when all methods are evaluated on those sites rather than on the PML V2 grid. If the advantage shrinks or vanishes, then part of the reported 0.836 RMSE comes from reconstructing the PML V2 product from its own driving variables.

Watch

Extended reading notes

Core claim

The paper's central claim is that encoding each location's spatiotemporal context as an implicit latent vector, instead of fitting local coefficients, lets one model capture spatial heterogeneity while still sharing statistical strength across the entire globe. The method predicts GPP and a set of interpretative weights simultaneously, with the prediction branch optimized by MSE and the explainability branch supervised by an auxiliary loss on a weighted linear combination of inputs. On the Climate2GPP test set the model reaches RMSE 0.836 and R2 0.932, beating LightGBM Large (RMSE 1.063), TabNet (RMSE 0.944), and spatial GWR (RMSE 1.937). On a few-shot comparison against GNNWR, the model shows a generalization gap of 0.073 versus 0.357, which the authors attribute to global learning rather than local fitting. The paper also exhibits spatial and temporal maps of temperature importance weights that shift across biomes and seasons, illustrating how the approach can expose where and when each driver matters.

Load-bearing premise

The paper treats the PML V2 GPP product as the ground-truth target even though PML V2 is itself a modeled estimate, and the ERA5 meteorological features are the same kind of variables that go into generating that product; if the target is essentially a deterministic function of the inputs, then the reported RMSE may measure how well the network reproduces the generating model rather than how well it predicts real vegetation growth.

Editorial extensions

If this is right

  • The same dual-branch design should transfer to other Earth-system variables whose drivers vary by region and season, such as evapotranspiration or soil moisture, provided a trustworthy target exists.
  • The per-location, per-time importance weights give spatial analysts a nonlinear replacement for GWR-style coefficient maps, useful for studying regional climate impacts.
  • Global shared-feature learning over the full dataset generalizes better to unseen times than purely local fitting, as indicated by the smaller generalization gap.
  • Linear attention and graph clustering keep the method tractable on tens of millions of samples, suggesting that it can scale to planetary-scale geoscience data.
  • The auxiliary loss ties interpretability to prediction, so the importance weights are optimized for the prediction task rather than derived post hoc.

Reading between the lines

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

  • If the PML V2 target is a function of the same ERA5 drivers, the absolute RMSE should not be read as an estimate of real-world carbon-flux predictive skill; the architecture claim still stands, but it needs independent validation.
  • The interpretative weights are correlation-based and task-optimized, not causal; adding interventions or synthetic counterfactuals would test whether the weights recover true driving mechanisms.
  • The STCG edges are fixed KNN constructs; learning the graph topology jointly with the node conditions could sharpen the local branch and may further improve the already small generalization gap.
  • The formulation resembles conditional prediction with latent context variables; ablating the condition branch versus the dual attention encoder would reveal how much of the gain comes from each component.
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 a dual-branch encoder-decoder neural network for modeling spatial and spatiotemporal heterogeneity in geographic regression tasks, with a case study predicting global gross primary productivity (GPP). The method combines a linear self-attention tabular encoder for location-invariant features with a spatiotemporal conditional graph (K-means clusters, KNN adjacency, GCN/LSTM aggregation) that produces an implicit condition vector, and a two-branch Transformer decoder that predicts the target and feature importance weights. The authors introduce the Climate2GPP dataset built from ERA5, MCD12C1, and PML V2 GPP, with about 50 million training samples (2001–2019) and 2.8 million test samples (2020). They report an RMSE of 0.836 and R² of 0.932 on the test set, outperforming LightGBM (1.063) and TabNet (0.944), and also compare against GWR, GTWR, and GNNWR on a sampled subset. Visualization of learned importance weights across regions and time is presented as interpretability evidence.

Significance. If the results are valid, the architecture would be a meaningful addition to spatial machine learning, combining global and local modeling in a single differentiable framework and scaling to tens of millions of samples. The paper's strengths are the scale of the dataset, the range of tabular and deep-learning baselines, and the attempt to provide spatial-temporal interpretable weights. However, the central claim that the model improves GPP prediction is undermined by the use of PML V2, a modeled product, as ground truth without independent validation; the reported margins are also from single runs with no error bars or code/data release. The interpretability claim is supported only by qualitative visualizations and is explicitly acknowledged by the authors as preliminary. The methodological innovations are plausible and empirically promising, but the evidence presented is not yet sufficient to establish that the method predicts real vegetation productivity rather than emulating the training target model.

major comments (5)
  1. [Section 4.1; Section 4.2] The regression target is PML V2 GPP (Zhang et al., 2019), a modeled product whose inputs include meteorological and remote-sensing data that overlap with the ERA5 features used here (Section 4.1). A low RMSE may therefore indicate that the model has learned to emulate the PML V2 generating process rather than to predict actual GPP. This concern is independent of the architecture. The paper does not validate against FLUXNET or any independent GPP product, so the reported RMSE of 0.836 and the margins over the baselines in Table 1 do not yet support the stated ecological claim. I recommend adding independent validation or substantially reframing the claim as PML V2 emulation.
  2. [Table 1; Section 4.2; Section 4.3] All reported metrics come from a single training run with no confidence intervals, and no code, data, or configuration files are provided. The differences between Ours (0.836), TabNet (0.944), and LightGBM Large (1.063) may be real, but stochastic training and random initialization can easily change these numbers. At minimum, the authors should run multiple seeds and report mean and standard deviation; ideally, they should release code and data to enable independent reproduction, especially given the 50-million-sample scale of the claimed benchmark.
  3. [Section 4.3; Tables 2 and 3] The comparison with GWR, GTWR, and GNNWR uses a different protocol from the main benchmark: 6,000 uniformly sampled grids, with training samples averaged over 19 years for GWR-series methods and separate weekly models for GWR/GNNWR. It is unclear how the proposed model is trained and tested on this subset, what its training data are, and why the reported Ours values in Table 2 differ from the full test set results in Table 1. The 'Generalization Gap' in Table 3 is computed from train and test RMSE values that may not be on the same data distribution, so the claim that global learning reduces overfitting is not established. Please clarify the exact experimental setup for Tables 2–3 and align the comparison protocol.
  4. [Section 3.2; Equations (5)–(8)] Several hyperparameters that materially affect the model are not specified: the number of clusters K, the number of neighbors k in the KNN adjacency matrix, the log-Gaussian kernel parameters σ and μ, the dimensions of the node embeddings, the number of GCN/LSTM layers, the Transformer decoder depth, and the number of epochs for the graph branch. Without these, the method is not reproducible, and the sensitivity of the reported results to K and k is unknown. Please provide a complete hyperparameter table or a configuration file.
  5. [Section 4.4; Section 6] The interpretability claim rests entirely on visual inspection of feature importance maps (Figures 4 and 5). There is no quantitative evaluation of whether these weights match known ecological drivers, no comparison with GWR coefficients, and no ablation or test of sensitivity to the auxiliary importance loss. The authors themselves state in Section 4.4 that 'this visualization analysis represents an initial exploration, and further work is needed to fully develop the interpretability of the model.' Given that interpretability is part of the title and contribution, this limitation should be addressed or the claim scaled back.
minor comments (6)
  1. [Section 3.2] In the paragraph after Equation (5), the text refers to 'Figure ??' with no figure number or label. Please fix the broken cross-reference.
  2. [Section 3.3] Equations (17) and (18) use the term 'intarget' where 'input target' or 'input' appears intended; please correct the terminology.
  3. [Table 2] The table caption is missing units for RMSE; it should state that RMSE is in g C m⁻² per 8-day period or otherwise define the units used in the dataset.
  4. [Section 4.3] The description of the 'few-shot dataset' is confusing: it says 6,000 grids are sampled, but then says training samples are 'the average data from every 8 days over 19 years.' Please clarify whether this means 19 yearly averages per grid or 8-day averages across the 19-year period, and how the test weeks (1, 10, 20, 30, 40) are derived from 2020 data.
  5. [Section 4.1] The dataset description says 'PML V2 GPP data aggregated every 8 days' but does not specify whether the GPP values are summed, averaged, or accumulated over the 8-day period; by comparison, the ERA5 radiation/evaporation/precipitation variables are explicitly stated to be summed. Please clarify the GPP aggregation to ensure the target units are interpretable.
  6. [References] Some references are incomplete or informal, e.g., 'Chen, J., et al., 2024. Excelformer: A neural network surpassing gbdts on tabular data' lacks venue or arXiv number; 'Vaswani, A., 2017. Attention is all you need. Advances in Neural Information Processing Systems' lists no volume or pages. Please complete the reference list.

Circularity Check

2 steps flagged · score 6.0 of 10

GPP prediction reduces to emulating the PML V2 model product, and the interpretative weights are fitted linear coefficients presented as explanations, making the central claims partly circular.

  1. other [Section 4.1, Dataset (Climate2GPP construction)]
    "To validate our approach, we created the Climate2GPP dataset, using the ERA5 climate dataset (Mu\~noz-Sabater et al., 2021), the MCD12C1.061 MODIS Land Cover dataset (Friedl and Sulla-Menashe, 2022), and the PML V2 0.1.7 GPP dataset (Zhang et al., 2019). From ERA5, we selected 26 climate parameters, with solar radiation, evaporation, and precipitation summed over 8-day periods, while the rest were averaged. GPP was similarly summed over 8-day intervals."

    The regression target, PML V2 GPP, is not measured GPP but the output of a coupled model that estimates GPP from meteorological forcings and remote sensing data. The predictor set is the same class of ERA5 meteorological forcings plus MODIS land cover. A model trained on these inputs to match PML V2 output is therefore largely learning to emulate the PML V2 generator, not to predict real vegetation GPP. The reported RMSE of 0.836 is thus an emulation score against a model product whose inputs overlap the predictors, so the claim of improved GPP prediction is circular with the dataset construction. The paper provides no independent validation against FLUXNET or another measured GPP product.

  2. fitted input called prediction [Section 3.3, Feature Importance Estimation Branch, Eq. (17)-(18)]
    "The predicted interpretable weights, denoted as wj, are then used to perform a weighted linear combination of the input intarget variables: \hat{y}_{interp} = \sum_{j=1}^{M} w_j x_j ... An auxiliary MSE loss is computed between \hat{y}_{interp} and the true target variable y to further refine the interpretability of this branch: L_interp = (1/N) \sum_i (y_i - \hat{y}_{interp,i})^2."

    The interpretative weights wj are optimized by directly minimizing the squared error between the linear combination of inputs and the target y. By construction, they are the coefficients of a linear surrogate fit to the target, not weights derived from the network's actual prediction mechanism. Presenting these fitted coefficients as 'interpretative weights' that reveal the 'dominant factors' of GPP is circular: the explanation is manufactured by the auxiliary loss, and the visualizations in Section 4.4 are not independent evidence of causal influence.

full rationale

The paper's central predictive architecture is trained on 2001-2019 and tested on 2020, so the comparison among methods is internally honest as an emulation benchmark. However, the ecological claim of predicting GPP is undermined by the choice of PML V2 as ground truth: PML V2 is itself a modeled product driven by meteorological and remote sensing inputs, and the paper's own features are ERA5 meteorology and MODIS land cover, i.e., the same class of physical drivers. Low RMSE therefore partly reflects reconstruction of the generating model rather than independent prediction of vegetation productivity. The paper's own Discussion states the model 'has so far been validated only on the Climate2GPP dataset,' confirming the absence of external validation. Separately, the interpretability branch fits weights to minimize a linear reconstruction error against the target, so calling them feature importance is a fitted quantity presented as an explanation. No load-bearing self-citation chain is present; the cited GNNWR work is by different authors. Overall, the architecture contribution may be real, but two central claims—predictive superiority for real GPP and interpretative discovery—reduce in part to the data-generation process and to the auxiliary fitting loss, warranting a partial circularity score of 6.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

All model weights and conditional embeddings are learned from the same Climate2GPP dataset, and the paper introduces no new physical entity. The free parameters are mostly unreported hyperparameters, and the main external assumptions concern the validity of PML V2 as ground truth and the sufficiency of ERA5 and land cover features.

free parameters (5)
  • K-means cluster count K = not reported
    Controls the granularity of shared spatiotemporal conditions; all grids in one cluster share a single condition vector.
  • KNN neighbor count k for adjacency matrix = not reported
    Defines connectivity in the spatiotemporal conditional graph and is chosen by hand.
  • Log-Gaussian kernel parameters sigma and mu = not reported
    Shape edge weights in Equation 8 and are chosen by hand without sensitivity analysis.
  • Network architecture hyperparameters = not reported
    Layer counts, hidden dimensions, embedding dimensions, and attention heads are not given, yet they determine model capacity and the reported RMSE.
  • Training schedule = batch 256, lr 1e-3 to 1e-4 after 10 epochs, 20 epochs
    Reported in Section 4.1 but no sensitivity analysis is provided for these choices.
assumptions (5)
  • domain assumption PML V2 GPP is a valid proxy for real gross primary productivity.
    This is the central benchmark assumption; if PML V2 is not effective ground truth, the reported RMSE does not measure prediction of real plant growth. Invoked in Section 4.1.
  • domain assumption ERA5 meteorological variables and MODIS land cover are sufficient inputs to predict GPP.
    The method uses these features without feature-selection or causal justification beyond correlation. Invoked in Section 4.1.
  • ad hoc to paper Locations within a K-means cluster share a single spatiotemporal condition vector.
    This reduces computational complexity but is not justified by an analysis of how cluster count affects the local-heterogeneity claim. Invoked in Section 3.2.
  • domain assumption The 2001-2019 to 2020 temporal split tests generalization under distribution shift.
    No analysis of climate stationarity or shift is provided, and 2020 is a single test year. Invoked in Section 4.1.
  • standard math Standard mathematical tools (K-means, GCN, LSTM, attention, node2vec, RoPE) are correct and applicable as used.
    The method relies on these off-the-shelf algorithms without proving new properties about them. Invoked throughout Section 3.
invented entities (1)
  • Implicit spatiotemporal conditional vector (node embedding in the Spatiotemporal Conditional Graph)
    purpose: Encodes location-specific time-varying conditions that modulate the global predictor.
    The vector is learned inside the model and has no external validation; its usefulness is shown only indirectly through test RMSE on the same benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Interpretable Implicit-Based Approach for Modeling Local Spatial Effects: A Case Study of Global Gross Primary Productivity." pith.science (2026). https://pith.science/paper/IHOEFUFW

@misc{pith2026250206170,
  author       = {Pith},
  title        = {Pith review of: An Interpretable Implicit-Based Approach for Modeling Local Spatial Effects: A Case Study of Global Gross Primary Productivity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IHOEFUFW}},
  note         = {Machine review of arXiv:2502.06170}
}
read the original abstract

In Earth sciences, unobserved factors exhibit non-stationary spatial distributions, causing the relationships between features and targets to display spatial heterogeneity. In geographic machine learning tasks, conventional statistical learning methods often struggle to capture spatial heterogeneity, leading to unsatisfactory prediction accuracy and unreliable interpretability. While approaches like Geographically Weighted Regression (GWR) capture local variations, they fall short of uncovering global patterns and tracking the continuous evolution of spatial heterogeneity. Motivated by this limitation, we propose a novel perspective - that is, simultaneously modeling common features across different locations alongside spatial differences using deep neural networks. The proposed method is a dual-branch neural network with an encoder-decoder structure. In the encoding stage, the method aggregates node information in a spatiotemporal conditional graph using GCN and LSTM, encoding location-specific spatiotemporal heterogeneity as an implicit conditional vector. Additionally, a self-attention-based encoder is used to extract location-invariant common features from the data. In the decoding stage, the approach employs a conditional generation strategy that predicts response variables and interpretative weights based on data features under spatiotemporal conditions. The approach is validated by predicting vegetation gross primary productivity (GPP) using global climate and land cover data from 2001 to 2020. Trained on 50 million samples and tested on 2.8 million, the proposed model achieves an RMSE of 0.836, outperforming LightGBM (1.063) and TabNet (0.944). Visualization analyses indicate that our method can reveal the distribution differences of the dominant factors of GPP across various times and locations.

Figures

Figures reproduced from arXiv: 2502.06170 by the authors.

Figure 1
Figure 1. The overall process of the method. encoder-decoder architecture. In the encoding stage, node in￾formation is aggregated in a spatiotemporal conditional graph using graph convolutional networks (GCN) and long short-term memory (LSTM) networks to encode location-specific spati￾otemporal heterogeneity as an implicit conditional vector. Ad￾ditionally, a self-attention-based encoder is utilized to extract location-invari… view at source ↗
Figure 2
Figure 2. The implementation details of the dual encoder and the dual decoder. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Spatiotemporal Conditional Graph Construction with Local Weight Sharing. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Temporal Variation of Feature Importance in Six Typical Regions. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Spatial Distribution of temperature 2m Importance Weights and GPP Prediction Results at 01, 10, 20, and 30. 4.4 Visualization Our visualization analysis aims to demonstrate the potential of our method in spatial analysis through three sets of results. The [PITH_FULL_I…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 22 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION list.label.output make.list.label ", " * write FUNCTION article output.bibitem list.label.output list.year output new.block title output new.block journal emphasize output format.volume.number output pages output new.block url output new.block note output fin.entry FUNCTION book output.bibitem list.l...

  2. [2]

    \" O ., Pfister, T., 2019

    Arik, S. \" O ., Pfister, T., 2019. TabNet: Attentive Interpretable Tabular Learning. CoRR , abs/1908.07442. http://arxiv.org/abs/1908.07442

    arXiv 2019

  3. [3]

    Scaling learning algorithms towards AI

    Bengio, Y., LeCun, Y., 2007. Scaling learning algorithms towards AI . Large Scale Kernel Machines, MIT Press

    work page 2007

  4. [4]

    Random forests

    Breiman, L., 2001. Random forests. Machine learning , 45, 5--32

    work page 2001

  5. [5]

    End-to-end object detection with transformers

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S., 2020. End-to-end object detection with transformers. European conference on computer vision, Springer, 213--229

    2020

  6. [6]

    Z., Wu, J., Sun, J., 2024

    Chen, J., Yan, J., Chen, Q., Chen, D. Z., Wu, J., Sun, J., 2024. Excelformer: A neural network surpassing gbdts on tabular data

    work page 2024

  7. [7]

    Xgboost: A scalable tree boosting system

    Chen, T., Guestrin, C., 2016. Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, 785--794

    work page 2016

  8. [8]

    Geographically neural network weighted regression for the accurate estimation of spatial non-stationarity

    Du, Z., Wang, Z., Wu, S., Zhang, F., Liu, R., 2020. Geographically neural network weighted regression for the accurate estimation of spatial non-stationarity. International Journal of Geographical Information Science , 34(7), 1353--1377

    work page 2020

Show all 30 references
  1. [9]

    Autogluon-tabular: Robust and accurate automl for structured data

    Erickson, N., Mueller, J., Shirkov, A., Zhang, H., Larroy, P., Li, M., Smola, A., 2020. Autogluon-tabular: Robust and accurate automl for structured data. arXiv preprint arXiv:2003.06505

    2020 arXiv

  2. [10]

    S., Brunsdon, C., Charlton, M., 2009

    Fotheringham, A. S., Brunsdon, C., Charlton, M., 2009. Geographically weighted regression. The Sage handbook of spatial analysis , 1, 243--254

    2009

  3. [11]

    S., Crespo, R., Yao, J., 2015

    Fotheringham, A. S., Crespo, R., Yao, J., 2015. Geographical and temporal weighted regression (GTWR). Geographical Analysis , 47(4), 431--452

    2015

  4. [12]

    Modis/terra+aqua land cover type yearly l3 global 0.05deg cmg v061

    Friedl, M., Sulla-Menashe, D., 2022. Modis/terra+aqua land cover type yearly l3 global 0.05deg cmg v061

    2022

  5. [13]

    Deep learning

    Goodfellow, I., Bengio, Y., Courville, A., Bengio, Y., 2016. Deep learning . 1, MIT Press

    2016

  6. [14]

    Revisiting Deep Learning Models for Tabular Data

    Gorishniy, Y., Rubachev, I., Khrulkov, V., Babenko, A., 2021. Revisiting Deep Learning Models for Tabular Data. CoRR , abs/2106.11959. https://arxiv.org/abs/2106.11959

    2021 arXiv

  7. [15]

    node2vec: Scalable feature learning for networks

    Grover, A., Leskovec, J., 2016. node2vec: Scalable feature learning for networks. Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, 855--864

    2016

  8. [16]

    E., Osindero, S., Teh, Y

    Hinton, G. E., Osindero, S., Teh, Y. W., 2006. A Fast Learning Algorithm for Deep Belief Nets. Neural Computation , 18, 1527--1554

    2006

  9. [17]

    PyTorch Frame: A Modular Framework for Multi-Modal Tabular Learning

    Hu, W., Yuan, Y., Zhang, Z., Nitta, A., Cao, K., Kocijan, V., Leskovec, J., Fey, M., 2024. PyTorch Frame: A Modular Framework for Multi-Modal Tabular Learning. arXiv preprint arXiv:2404.00776

    2024 arXiv

  10. [18]

    Transformers are RNN s: Fast autoregressive transformers with linear attention

    Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F., 2020. Transformers are RNN s: Fast autoregressive transformers with linear attention. H. D. III, A. Singh (eds), Proceedings of the 37th International Conference on Machine Learning, Proceedings of Machine Learning Research...

    2020

  11. [19]

    Lightgbm: A highly efficient gradient boosting decision tree

    Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., Liu, T.-Y., 2017. Lightgbm: A highly efficient gradient boosting decision tree. Advances in neural information processing systems , 30

    2017

  12. [20]

    N., Welling, M., 2016

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

    2016 arXiv

  13. [21]

    Global prediction of gross primary productivity under future climate change

    Lu, Q., Liu, H., Wei, L., Zhong, Y., Zhou, Z., 2024. Global prediction of gross primary productivity under future climate change. Science of The Total Environment , 912, 169239

    2024

  14. [22]

    et al., 2021

    Mu \ n oz-Sabater, J., Dutra, E., Agust \' -Panareda, A., Albergel, C., Arduini, G., Balsamo, G., Boussetta, S., Choulga, M., Harrigan, S., Hersbach, H. et al., 2021. ERA5-Land: A state-of-the-art global reanalysis dataset for land applications. Earth system science data , 13(...

    2021

  15. [23]

    V., Gulin, A., 2018

    Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V., Gulin, A., 2018. CatBoost: unbiased boosting with categorical features. Advances in neural information processing systems , 31

    2018

  16. [24]

    Global expansion of tropical cyclone precipitation footprint

    Qin, L., Zhu, L., Liu, B., Li, Z., Tian, Y., Mitchell, G., Shen, S., Xu, W., Chen, J., 2024. Global expansion of tropical cyclone precipitation footprint. Nature Communications , 15(1), 4824

    2024

  17. [25]

    Roformer: Enhanced transformer with rotary position embedding

    Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y., 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing , 568, 127063

    2024

  18. [26]

    Attention is all you need

    Vaswani, A., 2017. Attention is all you need. Advances in Neural Information Processing Systems

    2017

  19. [27]

    M., 2024

    Wang, S., Gao, K., Zhang, L., Yu, B., Easa, S. M., 2024. Geographically weighted machine learning for modeling spatial heterogeneity in traffic crash frequency and determinants in US. Accident Analysis & Prevention , 199, 107528

    2024

  20. [28]

    Geographically and temporally neural network weighted regression for modeling spatiotemporal non-stationary relationships

    Wu, S., Wang, Z., Du, Z., Huang, B., Zhang, F., Liu, R., 2021. Geographically and temporally neural network weighted regression for modeling spatiotemporal non-stationary relationships. International Journal of Geographical Information Science , 35(3), 582--608

    2021

  21. [29]

    Y., 2020

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., Philip, S. Y., 2020. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems , 32(1), 4--24

    2020

  22. [30]

    H., McVicar, T

    Zhang, Y., Kong, D., Gan, R., Chiew, F. H., McVicar, T. R., Zhang, Q., Yang, Y., 2019. Coupled estimation of 500 m and 8-day resolution global evapotranspiration and gross primary production in 2002--2017. Remote sensing of environment , 222, 165--182

    2019

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.