Pith. sign in

REVIEW 4 major objections 5 minor 22 references

Air in Your Neighborhood: Fine-Grained AQI Forecasting Using Mobile Sensor Data

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Deep learning on mobile-sensor data forecasts next-day Delhi AQI at 1 km² scale with 79% lower error than classical baselines.

desk verdict The first DL benchmark on AirDelhi is a useful starting point, but the headline 79% error reduction is not yet credible because the evaluation target is ambiguous and the abstract conflates MSE and RMSE. read the letter →

arxiv 2506.10332 v1 pith:QPE6UJ3I submitted 2025-06-12 cs.LG stat.ML

Aaryam Sharma This is my paper · ORCID
classification cs.LGstat.ML
keywords airqualityindexforecastingspatio-temporalgraphneuralnetworksmobilesensordataDelhidatasetPM2.5PM10inversedistanceweightingimputationgatedrecurrentunits
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 claims that deep sequence and graph models—GRUs, RNNs, and spatio-temporal graph networks (GCN and GAT)—can forecast next-day PM2.5 and PM10 concentrations at 1 km² resolution from the sparse, bus-mounted mobile-sensor AirDelhi dataset. The best models cut the paper's headline error metric from 90.6 (ridge) to 18.96 (GAT), a 79% reduction, with R² around 0.91, and the paper reports that these gains hold on coordinates never seen in training. It interprets this as evidence that AQI follows strong, learnable, short-term repetitive patterns and that spatial relations between neighboring grid cells change over time. If right, the work offers a path to localized air-quality alerts and targeted interventions without installing dense static sensor networks.

What carries the argument

The load-bearing mechanism is an Inverse Distance Weighting (IDW) imputer that converts sparse mobile readings into dense sequences every model trains on. It fills each empty 30-minute cell using the three closest coordinates, with distance power 3 and latitude/longitude over-weighted by 50× so that only immediate spatial neighbors at the same time of day from the previous two days contribute. On top of that, the paper builds temporal models (RNN, GRU) that read the imputed two-day hourly history, and spatio-temporal graph models (GCN/GAT followed by a GRU) that turn the grid into a k-nearest-neighbor graph (k=2–7); a ConvGRU treats the region as a 2D field to reach cells never sampled. The imputation is doing heavy lifting: the paper notes that in the ConvGRU representation only 4% of spatial nodes have true labels, and that STGNNs add only a small margin over plain GRUs because IDW already pools neighboring measurements.

What would settle it

Hold out genuinely measured readings from the AirDelhi grid, impute those cells with the paper's IDW rule as if they were unvisited, and score the trained models only on the actually-observed values at those cells; if the R² on these truly-unobserved measurements is near the baselines' 0.2 instead of the reported 0.91, the forecasting gains are artifacts of imputation.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a full deep-learning pipeline makes fine-grained AQI forecasting practical on data that was previously only analyzed with classical methods. The architecture combines an inverse-distance-weighting imputer with recurrent sequence models, optionally wrapped in GCN or GAT spatial layers and a 2D ConvGRU. The reported result is that the deep models outperform all classical baselines by a wide margin—R² jumps from roughly 0.19 for ridge, XGBoost, LightGBM, CatBoost, and IDW to 0.88–0.91 for RNN/GRU variants and up to 0.914 with GCN plus satellite images—and that performance on the extended test set of unseen coordinates is comparable or better for the recurrent models. The paper also claims two empirical insights about Delhi's AQI: there are strong repetitive short-term (two-day) temporal patterns, and the spatial influence of neighboring cells changes over time, which is why attention-based GAT improves with more neighbors while fixed-convolution GCN degrades.

Load-bearing premise

The results stand on the assumption that IDW imputation—using the three closest coordinates, over-weighted 50×, from the previous two days at the same time of day—produces a complete time series that faithfully represents true air quality without smoothing away the signal or leaking future information into training.

Editorial extensions

If this is right

  • Next-day AQI forecasts at 1 km² resolution are achievable from repeated mobile traversals, without installing a dense static sensor network.
  • Forecasts remain strong on coordinates never seen in training, suggesting the learned two-day temporal patterns are general and coordinate-agnostic.
  • The same models stay useful at forecast horizons up to 5 days, with R² dropping only 0.01–0.04 for recurrent and graph models.
  • Spatial relations between Delhi grid cells change over time: adding neighbors degrades fixed-convolution GCN but improves attention-based GAT, which can select informative neighbors.
  • Satellite image features add little for plain GRU/RNN models but reduce the STGNN accuracy drop on unseen nodes.

Reading between the lines

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

  • Inference: because plain GRUs (R² 0.893) almost match the best STGNNs (R² 0.911), most of the win over classical baselines likely comes from seeing a dense two-day hourly history rather than from graph structure; training a GRU on raw sparse sequences without IDW would test that directly.
  • Inference: the 'unseen coordinates' result is probably optimistic, since extended-set cells are still imputed from nearby training-phase cells; a stricter generalization test would exclude all imputed cells from evaluation.
  • Inference: the imputation-plus-recurrent pipeline is generic and could transfer to other urban mobile-sensing tasks—heat islands, traffic, noise—where the same routes are sampled repeatedly; the paper only suggests cross-city transfer as future work.
  • Inference: given the paper's own note that graph and 2D representations have fewer than 70 datapoints, the STGNN and ConvGRU numbers may be less stable than the GRU numbers; multiple training seeds would give error bars around the reported R².
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

4 major / 5 minor

Summary. The paper presents deep learning models (RNN, GRU, spatio-temporal GNNs, and a ConvGRU) for 24-hour-ahead AQI forecasting at 1 km^2 resolution using the AirDelhi mobile sensor dataset. The central claim is a large reduction in forecast error (from 90.6 to 18.96, described as a 79% MSE reduction) and successful generalization to unseen coordinates. The paper also reports ablation studies on model size, input sequence length, forecast horizon, and graph degree, from which it draws conclusions about repetitive short-term temporal patterns and changing spatial relations.

Significance. If the claims are substantiated, this would be a valuable first deep-learning study on the AirDelhi dataset with practical implications for fine-grained AQI forecasting from irregular mobile sensor data. Strengths include the release of code, the use of held-out test splits, and a fairly extensive set of ablations. However, the headline result is not currently credible because the evaluation metric is mislabeled (MSE versus RMSE) and, more importantly, because the paper does not establish whether the reported test metrics are computed on true sensor readings or on IDW-imputed targets. The significance of the claimed 79% improvement and the 'unseen coordinates' generalization therefore depends on unresolved evaluation details.

major comments (4)
  1. [§4.1, §4.5, Table 1] The paper never states whether the evaluation metrics in Table 1 are computed on genuine sensor readings or on the IDW-imputed grid described in Section 4.1. Section 4.5 notes that ConvGRU has only 4% true labels, and the 'Extended' evaluation set in Section 3 consists of coordinates with few readings and no training labels, so any target there must be imputed. If the test metrics are computed on imputed entries, a model whose input window contains the nearby-cell values used by the IDW rule can match the imputed target almost exactly without learning to forecast, while the baselines receive only sparse lagged features. This would explain the large performance gap and would mean the 'unseen coordinates' claim measures interpolation rather than forecast skill. Please clarify the evaluation protocol and re-report Table 1 on the subset of true ground-truth labels, including the fraction of true labels in each test set.
  2. [Abstract, Introduction, Table 1] The headline claim is numerically inconsistent with the table it cites. The abstract reports '71.654 MSE a 79% reduction' and the introduction states 'reducing MSE from 90.6 to 18.96'; both are consistent with the RMSE values in Table 1 (90.6 − 18.96 = 71.64), not with MSE, which would be on the order of 8208 to 359. Please correct the metric terminology throughout and report both MSE and RMSE consistently in the abstract, introduction, and Table 1.
  3. [§5, Table 1] The paper states that 'by doing a sweep on the number of RNN/GRU layers of each model, we obtain Table 1 of the best 24-hour forecast results,' which suggests that model selection was performed on the test set. Section 3 describes only an 80-20 train-test split and does not mention a validation set. Selecting hyperparameters on the test set optimistically biases the reported test performance. Please specify the validation procedure (e.g., a temporal validation split) and report the performance of models selected on validation data, or justify that the sweep did not use test labels.
  4. [§5.1, Figures 4–7] Section 5.1 states that 'for models using a Graph or 2D representation, the dataset size is fewer than 70 datapoints.' With such a small evaluation set, the R2 differences of 0.01–0.02 between models in Table 1 and the trends in the ablation figures are likely within noise. The paper provides no confidence intervals, standard deviations, or multiple-seed results. The 'new insights' (repetitive short-term patterns, changing spatial relations) are drawn primarily from these small-sample ablations; please provide uncertainty quantification or temper these claims accordingly.
minor comments (5)
  1. [§5] The text refers to 'columns titled Full Test' in Table 1, but the table columns are labeled 'Extended'; please align the terminology.
  2. [§5] The phrase 'both responders' should read 'both pollutants' (PM2.5 and PM10).
  3. [§4.2] For the baselines, the statement 'When any value is not available, the last reading is used' could introduce a forward-looking operation if the last reading is after the forecast origin; please clarify the temporal ordering of the carry-forward imputation.
  4. [§3] The sentence defining R2 writes 'R2 score is 1−RSS/TSS' without spacing or punctuation; please format the equation properly.
  5. [§5.1.3] Figure 6 is described as showing forecast-horizon sensitivity, but the text says models 'use satellite images' while Section 5.1 says only standard models are inspected for conciseness; please reconcile this inconsistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No demonstrated circularity; the IDW-imputation concern is a test-protocol ambiguity, not a by-construction equivalence.

full rationale

This paper is an empirical supervised-learning comparison, not a derivation: the central claim is a measured RMSE/R2 difference on the held-out AirDelhi test split (Table 1), with external classical baselines (XGB, Ridge, LightGBM, CatBoost, IDW) and standard deep architectures. No equation in the paper defines a target in terms of the model output, and no load-bearing result is justified by a self-citation: the author cites no prior work of his own, and the dataset and algorithms are external. The only step that could become circular is the IDW imputation used to complete the input grid (Section 4.1: 'Inverse Distance Weighting model, with the closest 3 points... only readings from the previous 2 days are used'). If Table 1's test targets were themselves the IDW-imputed values, then deep models would be rewarded for learning the imputation rule from nearby cells in their input window, and the comparison against lagged-feature baselines would not measure true forecast skill. However, the paper states in the Figure 2 caption that 'Any loss is backpropagated only over true ground truth labels,' and it never states that test metrics are computed on imputed entries; the Extended set is described as having sparse (<2/day) readings, which could still provide sparse true labels. Without an explicit statement that RMSE/R2 are computed on imputed targets, the concern is a data-integrity question to resolve by inspecting the released code, not a circularity demonstrable from the manuscript text. Accordingly, no circular step is established.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new theoretical entities, particles, forces, or conserved quantities. Its contribution is an empirical pipeline built from standard architectures and a domain-specific imputation scheme. The free parameters are hyperparameters and pre-processing choices that are swept or set by hand; the axioms are domain-level assumptions inherited from the AirDelhi dataset and the chosen spatial and temporal discretization.

free parameters (5)
  • IDW power and neighbor count = p=3, closest 3 points
    Set by hand in Section 4.1 for imputation; changing these changes the complete time series and all downstream model inputs.
  • lon/lat over-weight factor = 50x
    Section 4.1 states longitude and latitude are over-weighted by 50 times to keep imputation local; no sensitivity analysis is given.
  • k-nearest neighbors graph degree = 2 to 7 (swept)
    Section 4.4 builds graphs with k ranging from 2 to 7, and Figure 7 shows performance depends on k; the final choice is not pinned.
  • number of model layers and hidden size = swept, best around 3 or 6 layers
    Section 5.1.1 sweeps model size and selects the best based on test R2; final sizes are not reported in a fixed table.
  • input sequence length = swept, longer generally better
    Section 5.1.2 varies sequence length and reports R2 changes; the final sequence length for the main results is not stated explicitly.
assumptions (6)
  • domain assumption AirDelhi bus-mounted sensor readings accurately measure ambient PM2.5 and PM10 concentrations.
    The paper inherits the dataset from Chauhan et al. 2023 without independent validation of sensor accuracy or calibration.
  • domain assumption A 1 km^2 grid with 30-minute time bins is a faithful representation of neighborhood-level AQI.
    Adopted from Sahu et al. 2020 and Chauhan et al. 2023; no sensitivity analysis on grid size or time bin width is provided.
  • domain assumption Readings outside 5:30 AM to 10 PM IST can be discarded without systematic bias.
    Section 3 inherits this filter from Chauhan et al. 2023; nightly pollution patterns are not modeled or evaluated.
  • domain assumption IDW imputation using prior 2 days at the same time of day is leakage-free and preserves the true distribution.
    Section 4.1; the complete time series for all sequence models depends on this assumption being true.
  • domain assumption ESRI satellite image features capture population, foliage, and neighborhood type in a way useful for AQI prediction.
    Section 4.1 uses CNN features from ESRI images; there is no validation that these features carry predictive signal, and the paper notes they only help slightly.
  • standard math Standard GCN, GAT, GRU, and ConvGRU update equations behave as specified under the chosen training setup.
    Equations 2 through 4 are taken from prior literature; no formal verification is provided and no novel math is claimed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Air in Your Neighborhood: Fine-Grained AQI Forecasting Using Mobile Sensor Data." pith.science (2026). https://pith.science/paper/QPE6UJ3I

@misc{pith2026250610332,
  author       = {Pith},
  title        = {Pith review of: Air in Your Neighborhood: Fine-Grained AQI Forecasting Using Mobile Sensor Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QPE6UJ3I}},
  note         = {Machine review of arXiv:2506.10332}
}
read the original abstract

Air pollution has become a significant health risk in developing countries. While governments routinely publish air-quality index (AQI) data to track pollution, these values fail to capture the local reality, as sensors are often very sparse. In this paper, we address this gap by predicting AQI in 1 km^2 neighborhoods, using the example of AirDelhi dataset. Using Spatio-temporal GNNs we surpass existing works by 71.654 MSE a 79% reduction, even on unseen coordinates. New insights about AQI such as the existence of strong repetitive short-term patterns and changing spatial relations are also discovered. The code is available on GitHub.

Figures

Figures reproduced from arXiv: 2506.10332 by the authors.

Figure 1
Figure 1. Figures show the different nature of data. Bus routes are continuous paths, whereas typical [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of the feature engineering pipeline. IDW is used for imputation. Three primary [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the Convolutional GRU proposed in this paper. All satellite image models [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Test R2 score for PM2.5 and PM10 as model size changes. GRU has a stable performance, and ConvGRU starts improving, whereas other models begin overfitting. 6 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Test R2 score for PM2.5 and PM10 vs input sequence length. Generally, performance improves slightly as longer sequence is input, aside from a few outliers due to randomness in training 5.1.3 Forecast Horizon An important parameter to estimate is also the forecast horiz…
Figure 6
Figure 6. Figure 6: Test R2 score for PM2.5 and PM10 vs forecast horizon. Generally, performance decreases slightly for longer horizons, but the drop is not large [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Test R2 score for PM2.5 and PM10 vs node degree. While GCNs overfit with more neighbors, GATs improve performance by attending to correct neighbors. 6 Conclusion This report has presented the first in-depth modeling approach to forecast AQI over New Delhi at a fine-gra…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 15 canonical work pages

  1. [1]

    Air pollution in New Delhi is more severe than observed due to hygroscopicity-induced bias in aerosol sampling

    Ying Chen. Air pollution in New Delhi is more severe than observed due to hygroscopicity-induced bias in aerosol sampling. npj Clean Air, 1 0 (1): 0 1, March 2025. ISSN 3059-2240. doi:10.1038/s44407-024-00001-6. URL https://www.nature.com/articles/s44407-024-00001-6

    work page doi:10.1038/s44407-024-00001-6 2025

  2. [2]

    How is india trying to address air pollution?, 2024

    World Bank. How is india trying to address air pollution?, 2024. URL https://www.worldbank.org/en/country/india/publication/catalyzing-clean-air-in-india?

    work page 2024

  3. [3]

    Validation of Low - Cost Sensors in Measuring Real - Time PM10 Concentrations at Two Sites in Delhi National Capital Region

    Ravi Sahu, Kuldeep Kumar Dixit, Suneeti Mishra, Purushottam Kumar, Ashutosh Kumar Shukla, Ronak Sutaria, Shashi Tiwari, and Sachchida Nand Tripathi. Validation of Low - Cost Sensors in Measuring Real - Time PM10 Concentrations at Two Sites in Delhi National Capital Region . Sensors, 20 0 (5): 0 1347, February 2020. ISSN 1424-8220. doi:10.3390/s20051347. U...

    work page doi:10.3390/s20051347 2020

  4. [4]

    AirDelhi: Fine-Grained Spatio-Temporal Particulate Matter Dataset from Delhi for ML-Based Modeling

    Sachin Chauhan, Zeel Bharatkumar Patel, Sayan Ranu, Rijurekha Sen, and Nipun Batra. AirDelhi: Fine-Grained Spatio-Temporal Particulate Matter Dataset from Delhi for ML-Based Modeling . In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 75455--75468. Curran As...

    work page 2023

  5. [5]

    Predictive modeling of air quality in the Tehran megacity via deep learning techniques

    Abdullah Kaviani Rad, Mohammad Javad Nematollahi, Abbas Pak, and Mohammadreza Mahmoudi. Predictive modeling of air quality in the Tehran megacity via deep learning techniques. Scientific Reports, 15 0 (1): 0 1367, January 2025. ISSN 2045-2322. doi:10.1038/s41598-024-84550-6. URL https://www.nature.com/articles/s41598-024-84550-6

    work page doi:10.1038/s41598-024-84550-6 2025

  6. [6]

    Pruthi and Y

    D. Pruthi and Y. Liu. Low-cost nature-inspired deep learning system for PM2 .5 forecast over Delhi , India . Environment International, 166: 0 107373, August 2022. ISSN 01604120. doi:10.1016/j.envint.2022.107373. URL https://linkinghub.elsevier.com/retrieve/pii/S0160412022003002

    work page arXiv 2022

  7. [7]

    Alawi, Haslinda Mohamed Kamar, Ali Alsuwaiyan, and Zaher Mundher Yaseen

    Omer A. Alawi, Haslinda Mohamed Kamar, Ali Alsuwaiyan, and Zaher Mundher Yaseen. Temporal trends and predictive modeling of air pollutants in Delhi : a comparative study of artificial intelligence models. Scientific Reports, 14 0 (1): 0 30957, December 2024. ISSN 2045-2322. doi:10.1038/s41598-024-82117-z. URL https://www.nature.com/articles/s41598-024-82117-z

    work page doi:10.1038/s41598-024-82117-z 2024

  8. [8]

    Merayo, and Manuel Núñez

    Manuel Méndez, Mercedes G. Merayo, and Manuel Núñez. Machine learning algorithms to forecast air quality: a survey. Artificial Intelligence Review, 56 0 (9): 0 10031--10066, September 2023. ISSN 0269-2821, 1573-7462. doi:10.1007/s10462-023-10424-4. URL https://link.springer.com/10.1007/s10462-023-10424-4

    work page doi:10.1007/s10462-023-10424-4 2023

Show all 22 references
  1. [9]

    Sridhar, M Latha, Khalid Nazim Abdul Sattar, and G Manjula

    B D Parameshachari, G M Siddesh, V. Sridhar, M Latha, Khalid Nazim Abdul Sattar, and G Manjula. Prediction and Analysis of Air Quality Index using Machine Learning Algorithms . In 2022 IEEE International Conference on Data Science and Information System ( ICDSIS ) , pages 1--5...

    2022

  2. [10]

    AirPhyNet : Harnessing Physics - Guided Neural Networks for Air Quality Prediction , February 2024

    Kethmi Hirushini Hettige, Jiahao Ji, Shili Xiang, Cheng Long, Gao Cong, and Jingyuan Wang. AirPhyNet : Harnessing Physics - Guided Neural Networks for Air Quality Prediction , February 2024. URL http://arxiv.org/abs/2402.03784. arXiv:2402.03784 [cs]

    2024 arXiv

  3. [11]

    Messier, Brian J

    Yawen Guan, Margaret Johnson, Matthias Katzfuss, Elizabeth Mannshardt, Kyle P. Messier, Brian J. Reich, and Joon Jin Song. Fine-scale spatiotemporal air pollution analysis using mobile monitors on Google Street View vehicles. Journal of the American Statistical Association, 11...

    2020

  4. [12]

    ASTGC : Attention -based Spatio -temporal Fusion Graph Convolution Model for Fine -grained Air Quality Analysis

    Yaning Zhao, Shurui Fan, Kewen Xia, Yingmiao Jia, Li Wang, and Wenbiao Yang. ASTGC : Attention -based Spatio -temporal Fusion Graph Convolution Model for Fine -grained Air Quality Analysis . Air Quality, Atmosphere & Health, 16 0 (9): 0 1761--1775, September 2023. ISSN 1873-93...

    2023 doi

  5. [13]

    Predicting air quality via multimodal AI and satellite imagery

    Andrew Rowley and Oktay Karakuş. Predicting air quality via multimodal AI and satellite imagery. Remote Sensing of Environment, 293: 0 113609, August 2023. ISSN 00344257. doi:10.1016/j.rse.2023.113609. URL http://arxiv.org/abs/2211.00780. arXiv:2211.00780 [cs]

    2023

  6. [14]

    Deep- MAPS : Machine Learning based Mobile Air Pollution Sensing , March 2020

    Jun Song and Ke Han. Deep- MAPS : Machine Learning based Mobile Air Pollution Sensing , March 2020. URL http://arxiv.org/abs/1904.12303. arXiv:1904.12303 [cs]

    2020 arXiv

  7. [15]

    Anh Tuan Nguyen, Duy Hoang Pham, Bee Lan Oo, Yonghan Ahn, and Benson T. H. Lim. Predicting air quality index using attention hybrid deep learning and quantum-inspired particle swarm optimization. Journal of Big Data, 11 0 (1): 0 71, May 2024. ISSN 2196-1115. doi:10.1186/s40537...

    2024 doi

  8. [16]

    Optimized machine learning model for air quality index prediction in major cities in India

    Suresh Kumar Natarajan, Prakash Shanmurthy, Daniel Arockiam, Balamurugan Balusamy, and Shitharth Selvarajan. Optimized machine learning model for air quality index prediction in major cities in India . Scientific Reports, 14 0 (1): 0 6795, March 2024. ISSN 2045-2322. doi:10.10...

    2024 doi

  9. [17]

    Govt. India. Central pollution control board, Apr 2025. URL https://cpcb.nic.in/National-Air-Quality-Index/

    2025

  10. [18]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in P ython. Journal of Machine Learnin...

    2011

  11. [19]

    XGBoost : A Scalable Tree Boosting System

    Tianqi Chen and Carlos Guestrin. XGBoost : A Scalable Tree Boosting System . In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages 785--794, San Francisco California USA, August 2016. ACM. ISBN 978-1-4503-4232-2. doi:10.1...

    2016

  12. [20]

    Lightgbm: A highly efficient gradient boosting decision tree

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in ...

    2017

  13. [21]

    CatBoost : unbiased boosting with categorical features, January 2019

    Liudmila Prokhorenkova, Gleb Gusev, Aleksandr Vorobev, Anna Veronika Dorogush, and Andrey Gulin. CatBoost : unbiased boosting with categorical features, January 2019. URL http://arxiv.org/abs/1706.09516. arXiv:1706.09516 [cs]

    2019 arXiv

  14. [22]

    Graph Attention Networks , February 2018

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph Attention Networks , February 2018. URL http://arxiv.org/abs/1710.10903. arXiv:1710.10903 [stat]

    2018 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.