Pith. sign in

REVIEW 4 major objections 6 minor 44 references

LISA: Learning-Integrated Space Partitioning Framework for Traffic Accident Forecasting on Heterogeneous Spatiotemporal Data

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper proposes LISA, a framework that learns spatial partitions and forecasting models in one loop guided by prediction error, and reports average 13.0% improvements over four baselines on Iowa accident data.

desk verdict LISA's error-guided partition search is a real idea, but the sparse-cell confound and missing Spatial-Net baseline leave the main claim unsubstantiated. read the letter →

arxiv 2412.15365 v1 pith:LTU25E6D submitted 2024-12-19 cs.LG

Bang An , Xun Zhou , Amin Vahedian , Nick Street , Jinping Guan , Jun Luo This is my paper · ORCID
classification cs.LG
keywords trafficaccidentforecastingspatiotemporaldataminingspatialheterogeneityspacepartitioningensembleprediction-error-guidedLocalMoran'sIdeeplearning
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

Traffic accident patterns differ sharply across space: downtown and rural roads do not obey the same rules, so a single prediction model fit to the whole state is likely to miss local structure. The paper argues that the right response is not to pre-define regions with external knowledge but to learn the regions and the models together, with boundaries chosen so that validation prediction error keeps decreasing. It proposes LISA, a generic wrapper around any spatiotemporal deep network, that grows homogeneous partitions cell by cell and trains a separate model for each. On three years of Iowa accident data, LISA is reported to improve LSTM, ConvLSTM, DCRNN, and HintNet by an average of 13.0%, with the largest gains in the most heterogeneous areas.

What carries the argument

The load-bearing mechanism is a nested pair of procedures. I-HPT (Integrated Hierarchical Partitioning Training) repeatedly calls a Partition Learner on the remaining study area, peeling off one learned region at a time, and removes each completed partition from the space. Inside the Partition Learner, a Local Moran's I (or Geary's C) significance test chooses a seed set of spatially clustered high-accident cells, and an expansion loop compares the validation error of a single model trained on the union against the sum of errors of separate models for the seed and candidate pieces; the counter tolerates a number of failed expansions before stopping. The optional Spatial Gradient Search replaces naive candidate sampling with an error-gradient: it applies the current region's model to neighboring cells, smooths per-cell squared errors with a moving average of kernel size r, and takes the top-K lowest-error cells as the next candidate. This is what carries the claim: partition boundaries are produced by the same loss function that defines forecasting accuracy.

What would settle it

Run LISA on a synthetic grid with a dense-accident regime and a sparse-accident regime separated by a sharp boundary, then examine the learned partition boundary; if the boundary expands into the sparse regime merely because validation MSE there is trivially low, the error-guided stopping rule is not detecting homogeneity.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that the best partition of a heterogeneous region is the one that keeps validation error falling, so partition construction and model training should be one loop. LISA grows a seed region outward, at each step comparing a single model on the proposed larger region against two separate models on its pieces; expansion is accepted while the single model is not worse than the split, up to a tolerance. An optional Spatial Gradient Search ranks neighboring cells by the current model's smoothed per-cell squared error and keeps the K best, replacing arbitrary distance-based sampling. The paper reports that this self-guided procedure captures underlying heterogeneous patterns and improves LSTM, ConvLSTM, DCRNN, and HintNet by an average of 13.0% on the Iowa dataset, with the largest gains in the most heterogeneous regions.

Load-bearing premise

The expansion rule treats a low validation error of the current region's model on a candidate cell as evidence that the cell fits the same accident regime; sparse cells with near-zero MSE can make this signal misleading.

Editorial extensions

If this is right

  • Practitioners can drop manually designed region definitions: LISA outputs learned partitions and trained models together.
  • A model-agnostic wrapper means existing backbones such as LSTM, ConvLSTM, DCRNN, and HintNet can be upgraded without redesigning their architecture.
  • The larger the region and the more heterogeneous the accident patterns, the more the framework helps, matching the paper's reported 43.1% improvement for ConvLSTM and 16.4% for DCRNN on the most heterogeneous area.
  • The learned partitions are interpretable by construction: each one is a set of grid cells over which a single model was consistently better than splitting into two models.
  • Spatial correlation of predictions with true accident locations also improves under the Cross-K statistic, so the gains are not just lower average error but better placement of predicted events.

Reading between the lines

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

  • Editorial inference: the same 'a single model must beat two models' criterion could be applied to temporal or network partitions, not just spatial cells, since the stopping rule is stated purely in terms of validation error.
  • Editorial inference: the method's strongest reported gains appear in the largest 128x64 region, suggesting the biggest payoff is in statewide deployments where heterogeneity is unavoidable; city-scale applications may see smaller improvements.
  • Editorial inference: a direct test of the homogeneity assumption would be to compare learned partitions against independent feature-based clustering of the 47 input variables; if partitions and feature clusters diverge widely, the error criterion is selecting regions for a different reason than homogeneity.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes LISA, a learning-integrated space partitioning framework for traffic accident forecasting on heterogeneous spatiotemporal data. LISA alternates between training a base deep learning model (LSTM, ConvLSTM, DCRNN, or HintNet) and expanding a seed partition using either a basic neighbor-sampling scheme or a proposed Spatial Gradient Search that selects candidate cells where the current seed model has low validation MSE. The I-HPT algorithm iteratively forms partitions and trains a separate model per partition, using an error-based tolerance mechanism to decide when to stop expanding. Experiments on Iowa accident data (2016–2018) report that LISA improves each baseline network across four regions of increasing spatial heterogeneity, with an average improvement of 13.0%.

Significance. If the central claim holds, the idea of guiding space partitioning directly by prediction accuracy is a valuable and general contribution to spatiotemporal forecasting in heterogeneous domains. The framework is model-agnostic, the code is publicly available, and the evaluation on a real-world accident dataset with multiple baselines and feature ablations is a positive feature. However, the validity of the reported gains is not yet established because the Spatial Gradient Search and the termination rule appear to be confounded by the extreme sparsity of accident counts, and the experimental reporting lacks error bars and statistical tests. The omission of the directly relevant Spatial-Net baseline further weakens the comparison.

major comments (4)
  1. [§IV-C, Eq. (4), Algorithm 3] The Spatial Gradient Search selects candidate cells by taking the argmin over the smoothed per-cell MSE Loss_C computed with the seed model. On the Iowa dataset, the overwhelming majority of 5 km grid cells have zero accidents on most days, so any reasonable model predicts near zero in those cells and achieves near-zero MSE regardless of whether the cell's accident-generating process matches the seed region. Argmin over MSE thus preferentially selects sparse cells, not cells that are homogeneous in the sense of Definition 2. The ablation in Table III removes Spatial Gradient Search, but it does not control for the accident-count confound: a heuristic that selects the K lowest-accident candidate cells would likely also produce low validation MSE and might reproduce a large part of the gain. Without such a control, the reported improvements cannot be attributed to accuracy-guided partitioning.
  2. [§IV-B, Algorithm 2, Line 10] The termination condition lossnew > losss + lossc is not well defined dimensionally. If losss, lossc, and lossnew are average MSEs (as suggested by Eq. (4)), then the combined region's average is a weighted average of the seed and candidate averages, so the inequality can essentially never hold when the new model is reasonably trained; if they are total sums, then adding a large number of zero-accident cells artificially increases lossnew, biasing the algorithm toward terminating after sparse expansions. The manuscript does not specify which quantity is used, and the comparison is central to both the stopping rule and the claimed 'error-based' guidance. This issue also interacts with the sparse-cell confound of Algorithm 3.
  3. [§V-B, Table I] The central quantitative claim of a 13.0% average improvement is not supported by the reported experimental evidence. Single-run results without error bars, multiple seeds, or statistical tests are insufficient for deep learning models that the paper itself states are non-deterministic (see the tolerance discussion in §IV-B). Furthermore, the 13.0% average appears to be computed over percentage gains that are much larger for the weakest baselines (e.g., ConvLSTM at 43.1%) and small for the strongest baseline (e.g., HintNet at 3.5–6.3%); reporting the average of percentage improvements without bootstrapped confidence intervals or paired tests makes it impossible to assess whether the improvement over HintNet, the strongest baseline, is significant.
  4. [§II, §V-A.6] The related work section cites Spatial-Net [28] as 'a self-adaptive and model-agnostic deep learning framework for spatially heterogeneous datasets,' which is directly comparable to LISA's goal of automatically learning partitions. Spatial-Net is not included among the baselines in Table I or anywhere in the experiments, and no reason is given for its omission. Since Spatial-Net already addresses self-adaptive partitioning, the claim that LISA is the first to 'simultaneously learn partitions while training models' and the reported superiority over pre-defined partitions would be substantially more convincing if Spatial-Net were compared.
minor comments (6)
  1. [§V-A.2] The list of evaluation goals jumps from item (4) to item (5); item (4) appears to be missing.
  2. [§IV-A, Algorithm 1] The notation for the study area L is overloaded: L is used both as the spatial field and as a set, and later in the complexity analysis L is reused for the number of grid cells while n is used for both time steps and training cost per cell. This makes the complexity derivation hard to follow.
  3. [§IV-C, Algorithm 3] The parameter K in 'top K locations' is not listed among the framework inputs in Algorithm 1 or 2, and it only appears later in the parameter sensitivity study (§V-C.3). Please clarify how K is set during the main experiments.
  4. [Table I and Table III] Several baseline names contain formatting artifacts, e.g., 'H-C ONV LSTM', 'GSN ET', 'CONV LSTM', and 'H INT NET'; these should be corrected for readability.
  5. [References] Multiple references use abbreviated author lists such as 'B. et al.' (ref. [4]) and 'X. et al.' (refs. [27], [28]); full author lists should be provided in the final version.
  6. [§V-C.1, Table II] The feature-group ablation reports results only for LISA with HintNet; adding the original HintNet (without LISA) would help attribute the improvement to feature groups versus the partitioning framework.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: partition selection is tuned on validation data and the final claim is evaluated on held-out test data.

full rationale

The derivation chain is not circular. LISA learns partitions by optimizing validation loss in Algorithm 2 and via Spatial Gradient Search in Algorithm 3, but the reported 13.0% improvement is measured on the 2018 held-out test set after training on 2016–2017 data with a 20% validation split; the optimized quantity and the reported quantity are therefore not the same by construction. The seed initialization is the standard Local Moran’s I / Geary’s C statistics, and MSE is a standard external loss. Although several prior papers by overlapping authors are cited (HintNet [14], Hetero-ConvLSTM [15], Spatial-Net [28], ICDM/statistically-guided framework [27]), these are used as baselines and as supporting motivation for spatial ensembles, not as a proof of LISA’s central claim; the claim is tested against real data. The abbreviation “LISA test” in Algorithm 2 is explicitly explained as the Local Moran’s I test, so it is not a hidden redefinition. The potential tendency of Algorithm 3 to select sparse cells with trivially low MSE is a data-dependent risk in validation-error minimization, not a circular reduction: the final test predictions do not equal the inputs or fitted validation errors by definition.

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

The central claim rests on validation-error-guided expansion, which introduces several tuned hyperparameters (r, gamma, TopK, statistic choice) and three domain assumptions about the relationship between local error, spatial autocorrelation, and homogeneity. The paper itself notes model non-determinism but does not report variance.

free parameters (5)
  • radius r = 4
    Controls the Manhattan distance for sampling candidate neighbors; tuned on validation set in Fig. 7(b).
  • tolerance gamma = 1
    Number of consecutive failed expansions allowed before termination; tuned on validation set in Fig. 7(c).
  • TopK = 30
    Number of lowest-error candidate locations selected per expansion in Spatial Gradient Search; tuned on validation set in Fig. 7(a).
  • batch size threshold multiplier = not specified
    Controls the minimum number of candidate samples to train a model; no value given.
  • local spatial statistic choice = Moran's I vs Geary's C
    Two alternative statistics are tested; the better one is selected by validation performance.
assumptions (5)
  • domain assumption Local Moran's I and Geary's C can identify meaningful spatial clusters of accident counts that are homogeneous for prediction.
    Used to initialize seed partitions in Algorithm 2, Section IV-B.
  • domain assumption The validation error criterion lossnew > losss + lossc indicates that a merged region is too heterogeneous to be modeled by one network.
    Termination rule in Algorithm 2, lines 10-17.
  • domain assumption A model trained on the seed region generalizes to homogeneous neighboring cells, so low validation error on a candidate implies similarity.
    Basis of Spatial Gradient Search in Algorithm 3, Section IV-C.
  • domain assumption Accident records from Iowa DOT are complete, correctly geocoded, and representative of true accident locations.
    Data source described in Section III-A and used in all experiments.
  • standard math MSE is an appropriate loss for daily accident counts.
    Used in Eq. 4 and for all reported metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LISA: Learning-Integrated Space Partitioning Framework for Traffic Accident Forecasting on Heterogeneous Spatiotemporal Data." pith.science (2026). https://pith.science/paper/LTU25E6D

@misc{pith2026241215365,
  author       = {Pith},
  title        = {Pith review of: LISA: Learning-Integrated Space Partitioning Framework for Traffic Accident Forecasting on Heterogeneous Spatiotemporal Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LTU25E6D}},
  note         = {Machine review of arXiv:2412.15365}
}
read the original abstract

Traffic accident forecasting is an important task for intelligent transportation management and emergency response systems. However, this problem is challenging due to the spatial heterogeneity of the environment. Existing data-driven methods mostly focus on studying homogeneous areas with limited size (e.g. a single urban area such as New York City) and fail to handle the heterogeneous accident patterns over space at different scales. Recent advances (e.g. spatial ensemble) utilize pre-defined space partitions and learn multiple models to improve prediction accuracy. However, external knowledge is required to define proper space partitions before training models and pre-defined partitions may not necessarily reduce the heterogeneity. To address this issue, we propose a novel Learning-Integrated Space Partition Framework (LISA) to simultaneously learn partitions while training models, where the partitioning process and learning process are integrated in a way that partitioning is guided explicitly by prediction accuracy rather than other factors. Experiments using real-world datasets, demonstrate that our work can capture underlying heterogeneous patterns in a self-guided way and substantially improve baseline networks by an average of 13.0%.

Figures

Figures reproduced from arXiv: 2412.15365 by the authors.

Figure 1
Figure 1. Grid on the state of Iowa and mask map with road [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of Partition Leaner (PL) module. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The overall architecture of I-HPT (Brown box represents Partition Learner module). [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: An example of Spatial Gradient Search (The grey area [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Cross-K function measuring spatial correlation between [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 5
Figure 5. Figure 5: Four tested regions are indicated by colored boxes with [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Performance on different parameter settings D. Case Study We show an example of a successful prediction by LISA on Feb 5 th 2018 in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Case study on traffic accidents on Feb. 5 th , 2018. As observed, a majority of accidents and the number of crashes are correctly predicted in all regions. The blue circles highlight three major vehicle crash events that happened. In the selected urban area, severe mul…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 42 canonical work pages

  1. [28]

    Spatial-net: A self-adaptive and model-agnostic deep learning framework for spatially heterogeneous datasets,

    X. et al, “Spatial-net: A self-adaptive and model-agnostic deep learning framework for spatially heterogeneous datasets,” in Proceedings of the 29th International Conference on Advances in Geographic Information Systems, ser. SIGSPATIAL ’21, 2021

    work page 2021

  2. [1]

    2021 fatality data show increased traffic fatalities during pandemic,

    N. Media, “2021 fatality data show increased traffic fatalities during pandemic,” https://www.nhtsa.gov/press-releases/traffic-fatalities-estim ates-jan-sept-2021, Jun 2021

    work page 2021

  3. [2]

    Tennessee highway patrol predictive analytics,

    Gov, “Tennessee highway patrol predictive analytics,” 2017

    work page 2017

  4. [3]

    Self-supervised pre-training for robust and generic spatial-temporal representations,

    M. Hu, Z. Zhong, X. Zhang, Y . Li, Y . Xie, X. Jia, X. Zhou, and J. Luo, “Self-supervised pre-training for robust and generic spatial-temporal representations,” in 2023 IEEE International Conference on Data Mining (ICDM). IEEE, 2023, pp. 150–159

    work page 2023

  5. [4]

    Smoothing strategies combined with arima and neural net- works to improve the forecasting of traffic accidents,

    B. et al., “Smoothing strategies combined with arima and neural net- works to improve the forecasting of traffic accidents,” TheScientific- World, vol. 2014, pp. 152 375–12, 2014

    work page 2014

  6. [5]

    Referee-meta-learning for fast adaptation of locational fair- ness,

    C. et al., “Referee-meta-learning for fast adaptation of locational fair- ness,” arXiv.org, 2024

    work page 2024

  7. [6]

    Explaining the road accident risk: Weather effects,

    R. Bergel-Hayat, M. Debbarh, C. Antoniou, and G. Yannis, “Explaining the road accident risk: Weather effects,”Accident Analysis & Prevention, vol. 60, pp. 456–465, 2013

    work page 2013

  8. [7]

    A crash-prediction model for multilane roads,

    C. et al., “A crash-prediction model for multilane roads,” Accident Analysis & Prevention , vol. 39, no. 4, pp. 657–670, 2007

    work page 2007

Show all 44 references
  1. [8]

    Data mining of tree-based models to analyze freeway accident frequency,

    L.-Y . Chang and W.-C. Chen, “Data mining of tree-based models to analyze freeway accident frequency,” Journal of safety research, vol. 36, no. 4, pp. 365–375, 2005

    2005

  2. [9]

    A deep learning approach to the citywide traffic accident risk prediction,

    H. Ren, Y . Song, J. Wang, Y . Hu, and J. Lei, “A deep learning approach to the citywide traffic accident risk prediction,” 2018 21st International Conference on Intelligent Transportation Systems (ITSC) , pp. 3346– 3351, 2018

    2018

  3. [10]

    Combining satellite imagery and open data to map road safety,

    A. Najjar, S. Kaneko, and Y . Miyanaga, “Combining satellite imagery and open data to map road safety,” in Thirty-First AAAI Conference on Artificial Intelligence, 2017

    2017

  4. [11]

    Gsnet: Learning spatial-temporal correlations from geographical and semantic aspects for traffic accident risk forecasting,

    B. Wang, Y . Lin, S. Guo, and H. Wan, “Gsnet: Learning spatial-temporal correlations from geographical and semantic aspects for traffic accident risk forecasting,” in 2021 AAAI Conference on Artificial Intelligence (AAAI’21), 2021

    2021

  5. [12]

    Riskoracle: A minute- level citywide traffic accident forecasting framework,

    Z. Zhou, Y . Wang, X. Xie, L. Chen, and H. Liu, “Riskoracle: A minute- level citywide traffic accident forecasting framework,” 2020

    2020

  6. [13]

    Storm-gan+: spatio-temporal meta-gan for cross-city estimation of heterogeneous human mobility responses to covid-19,

    H. Bao, X. Zhou, Y . Xie, Y . Li, and X. Jia, “Storm-gan+: spatio-temporal meta-gan for cross-city estimation of heterogeneous human mobility responses to covid-19,” Knowledge and information systems , vol. 65, no. 11, pp. 4759–4795, 2023

    2023

  7. [14]

    Hintnet: Hierarchical knowledge transfer networks for traffic accident forecasting on heterogeneous spatio-temporal data,

    A. et al., “Hintnet: Hierarchical knowledge transfer networks for traffic accident forecasting on heterogeneous spatio-temporal data,” Proceed- ings of the 2022 SIAM International Conference on Data Mining , 2022

    2022

  8. [15]

    Hetero-convlstm: A deep learning approach to traffic accident prediction on heterogeneous spatio-temporal data,

    Z. Yuan, X. Zhou, and T. Yang, “Hetero-convlstm: A deep learning approach to traffic accident prediction on heterogeneous spatio-temporal data,” in Proceedings of the 24th ACM SIGKDD International Confer- ence on Knowledge Discovery & Data Mining , 2018, pp. 984–992

    2018

  9. [16]

    Traffic accident analysis using decision trees and neural networks,

    M. M. Chong, A. Abraham, and M. Paprzycki, “Traffic accident analysis using decision trees and neural networks,” ArXiv, 2004

    2004

  10. [17]

    A novel variable selection method based on frequent pattern tree for real-time traffic accident risk prediction,

    L. Lin, Q. Wang, and A. W. Sadek, “A novel variable selection method based on frequent pattern tree for real-time traffic accident risk prediction,” Transportation Research Part C, vol. 55, 2015

    2015

  11. [18]

    Analysis of traffic accident severity using decision rules via decision trees,

    J. Abell ´an, G. L ´opez, and J. de O ˜na, “Analysis of traffic accident severity using decision rules via decision trees,” Expert Systems with Applications, vol. 40, no. 15, pp. 6047–6054, 2013

    2013

  12. [19]

    Modeling traffic accident occurrence and involvement,

    M. A. Abdel-Aty and A. Radwan, “Modeling traffic accident occurrence and involvement,” Accident Analysis & Prevention , vol. 32, no. 5, pp. 633–642, 2000

    2000

  13. [20]

    Pdformer: Propagation delay-aware dynamic long-range trans- former for traffic flow prediction,

    J. et al., “Pdformer: Propagation delay-aware dynamic long-range trans- former for traffic flow prediction,” arXiv.org, 2024

    2024

  14. [21]

    St-ifgsm: Enhancing robustness of human mobility signature identification model via spatial- temporal iterative fgsm,

    M. Hu, X. Zhang, Y . Li, X. Zhou, and J. Luo, “St-ifgsm: Enhancing robustness of human mobility signature identification model via spatial- temporal iterative fgsm,” in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . New York, NY , USA: A...

    2023

  15. [22]

    Spatialrank: urban event ranking with ndcg optimization on spatiotemporal data,

    A. et al., “Spatialrank: urban event ranking with ndcg optimization on spatiotemporal data,” in Proceedings of the 37th International Confer- ence on Neural Information Processing Systems , ser. NIPS ’23. Red Hook, NY , USA: Curran Associates Inc., 2024

    2024

  16. [23]

    Urban traffic dynam- ics prediction—a continuous spatial-temporal meta-learning approach,

    Y . Zhang, Y . Li, X. Zhou, J. Luo, and Z.-L. Zhang, “Urban traffic dynam- ics prediction—a continuous spatial-temporal meta-learning approach,” ACM Trans. Intell. Syst. Technol. , 2022

    2022

  17. [24]

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting,

    Y . Li, R. Yu, C. Shahabi, and Y . Liu, “Diffusion convolutional recurrent neural network: Data-driven traffic forecasting,” 2018

    2018

  18. [25]

    Sdcae: Stack denoising convolutional autoencoder model for accident risk prediction via traffic big data,

    C. Chen, X. Fan, C. Zheng, L. Xiao, M. Cheng, and C. Wang, “Sdcae: Stack denoising convolutional autoencoder model for accident risk prediction via traffic big data,” 2018 Sixth International Conference on Advanced Cloud and Big Data (CBD) , pp. 328–333, 2018

    2018

  19. [26]

    Foresee urban sparse traffic accidents: A spatiotemporal multi-granularity perspective,

    Z. Zhou, Y . Wang, X. Xie, L. Chen, and C. Zhu, “Foresee urban sparse traffic accidents: A spatiotemporal multi-granularity perspective,” IEEE Transactions on Knowledge and Data Engineering , pp. 1–1, 2020

    2020

  20. [27]

    A statistically-guided deep network transformation and moderation framework for data with spatial heterogeneity,

    X. et al., “A statistically-guided deep network transformation and moderation framework for data with spatial heterogeneity,” in2021 IEEE International Conference on Data Mining (ICDM) , 2021, pp. 767–776

    2021

  21. [29]

    Iowa crash analysis tool,

    “Iowa crash analysis tool,” 2022. [Online]. Available: https://icat.iowad ot.gov/#

    2022

  22. [30]

    [Online]

    IowaDot, 2022. [Online]. Available: https://mesonet.agron.iastate.edu/ RWIS/

    2022

  23. [31]

    MAP, 2022

    H. MAP, 2022. [Online]. Available: https://developer.here.com/docum entation/places/dev guide/topics/categories.html

    2022

  24. [32]

    Spatio-temporal data mining: A survey of problems and methods,

    G. Atluri, A. Karpatne, and V . Kumar, “Spatio-temporal data mining: A survey of problems and methods,” ACM Comput. Surv. , vol. 51, no. 4, Aug 2018

    2018

  25. [33]

    Local indicators of spatial association,

    L. Anselin, “Local indicators of spatial association,” Geographical Analysis, vol. 27, no. 2, pp. 93–115, 1995

    1995

  26. [34]

    A comparison of spatial heterogeneity with local cluster detection methods for chronic respiratory diseases,

    W. et al., “A comparison of spatial heterogeneity with local cluster detection methods for chronic respiratory diseases,” F1000 research , vol. 6, p. 1819, 2018

    2018

  27. [35]

    P. M. Dixon, Ripley’s K Function . John Wiley & Sons, Ltd, 2014

    2014

  28. [36]

    Adam: A method for stochastic optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization.”

  29. [37]

    Pytorch: An imperative style, high-performance deep learning library,

    P. et al., “Pytorch: An imperative style, high-performance deep learning library,” in Advances in Neural Information Processing Systems 32 . Curran Associates, Inc., 2019

    2019

  30. [38]

    D. C. Montgomery, Introduction to linear regression analysis , fifth edition.. ed., ser. Wiley series in probability and statistics. Hoboken, New Jersey: John Wiley & Sons Ltd, 2012

    2012

  31. [39]

    Classification and regression trees,

    J. V . Ryzin, L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone, “Classification and regression trees,” Journal of the American Statistical Association, vol. 81, no. 393, pp. 253–253, 1986

    1986

  32. [40]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  33. [41]

    Convolutional lstm network: A machine learning approach for precipitation nowcasting,

    X. Shi, Z. Chen, H. Wang, D.-Y . Yeung, W.-k. Wong, and W.-c. Woo, “Convolutional lstm network: A machine learning approach for precipitation nowcasting,” 2015

    2015

  34. [42]

    Flow cross k-function: a bivariate flow analytical method,

    R. Tao and J.-C. Thill, “Flow cross k-function: a bivariate flow analytical method,” International journal of geographical information science : IJGIS, vol. 33, no. 10, pp. 2055–2071, 2019

    2019

  35. [43]

    Ninth weather-related traffic death in iowa,

    D. M. R. Media, “Ninth weather-related traffic death in iowa,” 2018. [Online]. Available: https://www.desmoinesregister.com/story/news/201 8/02/07/ninth-weather-related-traffic-death-reported-iowa/314765002/

    2018

  36. [44]

    K.-T. N. Media, Feb 2018. [Online]. Available: https://www.kcrg.com /content/news/Update-I-380-NB-delayed-due-to-serious-injury-crash-4 72790753.html

    2018

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.