ALCo-FM: Adaptive Long-Context Foundation Model for Accident Prediction

Ahmad Mohammadshirazi; Pinaki Prasad Guha Neogi; Rajiv Ramnath

REVIEW 4 major objections 6 minor 34 references

ALCo-FM: Adaptive Long-Context Foundation Model for Accident Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A volatility-gated multimodal transformer predicts city-scale accident risk at 0.94 accuracy and 0.92 F1 across 15 U.S. cities, and transfers to new cities after light fine-tuning.

desk verdict A genuinely new adaptive-context architecture, but the reported SOTA numbers rest on an undefined data split and a 1h/3h window contradiction that must be fixed before the results can be trusted. read the letter →

arxiv 2507.08153 v1 pith:GZDI7AT6 submitted 2025-07-10 cs.LG cs.AI

Pinaki Prasad Guha Neogi , Ahmad Mohammadshirazi , Rajiv Ramnath This is my paper · ORCID

classification cs.LGcs.AI

keywords accidentpredictionlong-contexttransformermultimodalfusionH3hexagonalgridadaptivecontextwindowuncertaintycalibrationspatio-temporalgraphurbanriskforecasting

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

ALCo-FM sets out to show that rare, high-impact events like traffic accidents can be forecast at city scale by a single model that adapts how much history it uses, fuses multiple data types, and reasons over space as well as time. The model computes a volatility score from its own embeddings and uses it to choose a 1-, 3-, or 6-hour look-back per region; it then fuses numerical time series with map imagery, propagates information across a hexagonal H3 grid, and reports confidence through Monte Carlo dropout. Trained on 1,771 regions in 15 U.S. cities and fine-tuned on three held-out cities, it reports 0.94 accuracy, 0.92 F1, and 0.04 expected calibration error, beating more than 20 baselines. The paper's claim is that adaptive long-context multimodal reasoning, not a single architectural trick, is what moves rare-event risk prediction.

What carries the argument

The load-bearing mechanism is the volatility pre-score $u=(\sigma_{\mathrm{num}}+\sigma_{\mathrm{vis}})/2$, where $\sigma$ is the mean standard deviation across the temporal or visual token dimensions of the current hour's dual embedding. Thresholds at the 33rd and 67th percentiles of the training distribution map $u$ to a look-back window $w \in \{1,3,6\}$ hours, so volatile cells get more history and stable cells save compute. Around that gate, the architecture stacks four modules that each add a distinct capability: a continuous-time transformer encodes the hourly numerical time series, a vision transformer encodes 256x256 map tiles, shallow bidirectional cross-attention fuses the two token streams, and a GAT layer plus BigBird-style sparse global attention propagate information across the H3 hexagonal graph. Monte Carlo dropout with ten stochastic passes produces the final mean prediction and an uncertainty interval used for calibration.

What would settle it

Re-run ALCo-FM with a strict chronological split, training on earlier months and testing on strictly later months with a gap between the two periods, and compare F1 and ECE against the reported 0.92 and 0.04; a material drop under that split would mean the central claim as reported does not follow. The reproduction should also give the baselines matched adaptive windows and the same visual input.

Watch

Extended reading notes

Core claim

The paper's central claim is that accident risk at urban scale is better predicted by a foundation-style model that treats context length as an input-dependent decision rather than a fixed hyperparameter. As evidence it presents ALCo-FM: paired encoders produce embeddings for each hour of numerical and visual data; a volatility pre-score summarizes how much those embeddings fluctuate and gates the history window to 1, 3, or 6 hours; two cross-attention layers let the modalities read each other; a GAT layer passes messages between neighboring H3 hexagons; a BigBird-style sparse transformer mixes those local signals with shared city-wide tokens; and Monte Carlo dropout turns the final MLP prediction into a calibrated risk with an uncertainty interval. In the reported experiments this stack reaches 0.94 accuracy and 0.92 F1 on 1,771 regions, with ECE 0.04, and after only five epochs of fine-tuning on the final GAT and MLP layers, it holds F1 near 0.90 on Columbus, Portland, and Oklahoma City.

Load-bearing premise

The reported numbers stand on the unstated assumption that the train/validation/test split respects time, since a random split would let the model see hours immediately before or after a test accident and inflate F1 in a rare-event time series, and the baselines are also compared with a fixed three-hour window and no visual input, so the margin over them may partly reflect a handicap rather than a fair comparison.

Editorial extensions

If this is right

Adaptive gating is the highest-value component: adding it after MC-dropout calibration raises F1 from 0.88 to 0.92 and cuts ECE from 0.05 to 0.04.
Sparse global attention lets the model reason about city-wide conditions without paying quadratic attention over all 1,771 cells.
Monte Carlo dropout gives each prediction a confidence interval, which matters when emergency dispatch decisions carry asymmetric costs.
With only the final GAT and MLP layers fine-tuned for five epochs, the pretrained representations reach F1 around 0.89 to 0.90 on held-out cities.
Under severe class imbalance, accuracy is less informative than F1, so the model's precision-recall balance is the performance claim to watch.

Reading between the lines

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

If the temporal split is valid, the same volatility-gated recipe is a natural candidate for other rare-event spatio-temporal forecasting problems, such as flash flooding, power outages, or disease outbreaks, where context length should grow when conditions become unstable.
The threshold-based gate could be replaced by a learned or per-city adaptive policy, and testing that variant would show whether the simple percentiles are a limitation or a feature.
Because the reported baselines were run without the visual modality, an ablation that gives them the same map tiles would isolate how much of the margin comes from multimodal fusion rather than from the rest of the architecture.

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.

Desk Editor's Note

read the letter

Two things to know before you read this one. First, the core idea—a volatility pre-score that gates how many hours of history to attend to, computed from both numerical and visual token streams—is genuinely new in the accident-prediction space, and the full pipeline (ContiFormer + T2T-ViT, cross-attention fusion, local GAT, BigBird global attention, MC dropout) is assembled in a way I haven't seen before. The dataset construction is careful, with sensible imputation and feature-selection analysis in the appendix, and they ship code and data. Second, the evaluation as written does not support the headline claims. There is no description of the train/validation/test split anywhere, and with 1–6 hour look-back windows over rare, temporally clustered events, a random row split could put the precursors or aftermath of test accidents in training. That alone makes the F1=0.92 and ECE=0.04 unverifiable. The authors also claim a 'fair comparison' with 20+ baselines while giving those baselines only a fixed 3-hour window and no visual context—that is an ablation, not a SOTA comparison. And there is an internal contradiction: Section 3.2 and Section 4.1 define 1-hour atomic windows, but Appendix C.5 says the data is grouped into 3-hour intervals. The label definition and the prediction unit depend on which is right. The progressive ablation in Table 2 is the most informative experiment; it shows each component adds something, with adaptive gating giving the largest F1 jump from 0.88 to 0.92. That internal consistency is a point in the paper's favor. I'd also want error bars or multiple seeds, since MC dropout is right there. The free parameters (tau_low, tau_high, number of global tokens) are tuned on validation data, which is normal and not a circularity problem. Citation pattern looks fine. If the split is chronological and the time-window discrepancy is resolved, this could be a solid contribution to accident prediction. But right now the central empirical claim is not well-supported. My recommendation: send it to peer review, because the architecture deserves scrutiny and the evaluation flaws are fixable. A referee round will either force the authors to clarify the split or will catch a real leakage problem. Either way, the field learns something.

Referee Report

4 major / 6 minor

Summary. ALCo-FM is an accident-risk prediction model that combines continuous-time and vision encoders with a volatility-based adaptive context window, cross-modal attention, a local GAT layer, and BigBird-style sparse global attention over H3 hexagonal cells, plus Monte Carlo dropout for calibration. The paper reports F1=0.92, accuracy=0.94, and ECE=0.04 on 1,771 regions across 15 U.S. cities, claims to outperform 20+ baselines, and reports transfer results on three held-out cities after minimal fine-tuning. The main claimed contributions are the adaptive windowing mechanism and the unified multimodal architecture for rare-event risk forecasting.

Significance. If the empirical claims are validated, ALCo-FM would be a useful multimodal, calibration-aware approach to rare-event urban risk prediction, and the release of code and data is a concrete reproducibility asset. The architecture is clearly described and the progressive ablation is a strength. However, the central SOTA claim is currently under-supported because the evaluation protocol is incomplete: the train/validation/test split is not defined, the temporal granularity is internally inconsistent, and baselines are restricted to a setting that is not comparable to the full ALCo-FM. These issues need to be resolved before the reported numbers can be taken as evidence of state-of-the-art performance.

major comments (4)

[Section 5.1 / Table 3] The manuscript never specifies how the data are partitioned into training, validation, and test sets. Since the prediction unit is an (H3 cell, time slice) and the model uses up to 6 hours of prior context, a random row split would allow training samples to contain the immediate precursors or aftermaths of test-slice accidents, artificially inflating F1 and ECE and making the comparison to baselines uninterpretable. Please state the split (ideally chronological on the temporal axis), the temporal gap between the last training timestamp and the first test timestamp, and report all main-table metrics on the test set only.
[Section 3.2 vs Appendix C.5] Section 3.2 defines the input as uniform 1-hour atomic windows, while Appendix C.5 states that the dataset is grouped into 3-hour intervals for label definition, weather aggregation, and temporal feature extraction. These two descriptions are incompatible. Resolve this by specifying the actual temporal resolution of the label and input units, and clarify what w = 1, 3, 6 means in each case, because the leakage argument and the adaptive-window benefit both depend on this.
[Section 5.3] The baselines are deliberately restricted to a fixed 3-hour window and are not provided the visual modality, while ALCo-FM uses adaptive 1/3/6-hour windows and map imagery. This is not an apples-to-apples comparison and limits the validity of the "outperforming 20+ baselines" claim. Either provide baselines with comparable context lengths and input modalities, or report ALCo-FM's performance in the same 3-hour, numeric-only setting.
[Tables 2-4] No error bars, confidence intervals, or numbers of seeds are reported, and the claimed improvements in F1 over the strongest baselines are small (e.g., 0.92 vs 0.85 in Table 3) and could lie within run-to-run variance. Please report mean and standard deviation over multiple seeds (or an equivalent significance analysis) for the main results and the ablation.

minor comments (6)

[Section 5.3, first paragraph] There is a typo: "we reportAccu-racy" should be "we report Accuracy".
[Section 4.4] The number of global tokens G inserted into the sparse attention block is never specified; please state the value used in the experiments.
[Section 4.1] The thresholds tau_low and tau_high are described as set to the 33rd/67th percentiles and then refined by validation grid-search, but the grid range and the criterion for the final selection are not given; please clarify whether the validation set used for this tuning is temporally disjoint from the training set.
[Section 4.5] The text says dropout layers remain active during inference with p = 0.2, but it is not clear where the dropout is applied in the 2-layer MLP head or whether it is also applied in the encoders; please specify the dropout locations.
[Section 5.4] The held-out city fine-tuning uses early stopping based on validation F1, but the construction of that validation set within each held-out city is not described; please specify its temporal structure and size.
[Section 5.3, Table 3] The phrase "the complete dataset spanning 15 U.S. cities" is ambiguous; clarify whether the reported numbers are computed on the full dataset, a held-out test split, or both, and report the corresponding split sizes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ALCo-FM is a standard empirical pipeline whose reported numbers rest on external data and baselines, not on a self-referential reduction.

full rationale

ALCo-FM's derivation chain is an empirical pipeline: assemble heterogeneous data on H3 cells, encode numeric and visual inputs, set volatility thresholds on training/validation data to choose a 1/3/6-hour context window, fuse modalities, propagate spatially with GAT and sparse global attention, and classify with MC-dropout. The threshold tuning in Section 4.1 is hyperparameter selection; the reported risk score is the output of a learned classification head, so no reported metric is the same quantity as the fitted thresholds. Baseline comparisons in Section 5.3 are against external published methods, and the held-out-city experiments in Section 5.4 use cities excluded from pretraining. The self-citations present in the paper, such as the Moosavi et al. 2019 dataset, Karimi Monsefi et al. 2023 and Monsefi et al. 2022 for map-image motivation, Mohammadshirazi et al. 2024 as analogies, and Guha Neogi & Goswami 2021 mentioned as future work, are not load-bearing for any mathematical claim; no uniqueness theorem or ansatz is imported from the authors' prior work. The missing statement of the train/validation/test split, the fixed 3-hour no-vision configuration of baselines, and the inconsistency between 1-hour atomic windows and 3-hour interval grouping are serious evaluation-validity risks, but they are not circularity: they do not make any prediction equivalent to its input by construction. No Eq. X = Eq. Y reduction, fitted-parameter-renamed-as-prediction, or self-citation chain is present, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the correctness of the assembled dataset, a valid temporal split, and fair baseline comparisons. The adaptive gating thresholds (tau_low, tau_high) are fitted to validation data, and the global token count is unreported. No new physical or conceptual entities are introduced.

free parameters (3)

tau_low = 33rd percentile of training-set volatility u, refined by validation grid search
Threshold determining whether the model uses 1h, 3h or 6h context; tuned on validation to optimize F1 and ECE (Section 4.1).
tau_high = 67th percentile of training-set volatility u, refined by validation grid search
Upper threshold for adaptive window selection (Section 4.1).
Global token count G = not reported
The BigBird-style sparse attention inserts G learnable global tokens (Section 4.4); the paper does not report G, so it is an unstated hyperparameter.

assumptions (4)

domain assumption The accident, weather, demographic, and map data are accurately aligned and representative.
The dataset joins Moosavi et al. accidents, ZIP-based demographics, ASOS weather, and OSM tiles on H3 cells (Section 3). Errors or misalignment would invalidate all reported metrics.
domain assumption The train/validation/test split is temporally valid.
The paper does not describe the split; if it is random, temporal leakage inflates F1 in rare-event prediction.
domain assumption Baseline models were implemented and tuned fairly.
Baselines are given a fixed 3-hour window without visual context; if under-tuned, the comparison overstates ALCo-FM's gains.
standard math Standard transformer and GAT machinery (softmax attention, layer norm, residual connections) behaves as assumed.
Sections 4.2-4.4 rely on conventional attention and graph attention primitives.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ALCo-FM: Adaptive Long-Context Foundation Model for Accident Prediction." pith.science (2026). https://pith.science/paper/GZDI7AT6

@misc{pith2026250708153,
  author       = {Pith},
  title        = {Pith review of: ALCo-FM: Adaptive Long-Context Foundation Model for Accident Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZDI7AT6}},
  note         = {Machine review of arXiv:2507.08153}
}

read the original abstract

Traffic accidents are rare, yet high-impact events that require long-context multimodal reasoning for accurate risk forecasting. In this paper, we introduce ALCo-FM, a unified adaptive long-context foundation model that computes a volatility pre-score to dynamically select context windows for input data and encodes and fuses these multimodal data via shallow cross attention. Following a local GAT layer and a BigBird-style sparse global transformer over H3 hexagonal grids, coupled with Monte Carlo dropout for confidence, the model yields superior, well-calibrated predictions. Trained on data from 15 US cities with a class-weighted loss to counter label imbalance, and fine-tuned with minimal data on held-out cities, ALCo-FM achieves 0.94 accuracy, 0.92 F1, and an ECE of 0.04, outperforming more than 20 state-of-the-art baselines in large-scale urban risk prediction. Code and dataset are available at: https://github.com/PinakiPrasad12/ALCo-FM

Figures

Figures reproduced from arXiv: 2507.08153 by the authors.

**Figure 1.** Deaths due to road accidents in various developed countries (2000–2022). former for efficient, city-wide context aggregation across all H3 cells (Uber Technologies, 2023). • Foundation-Scale Calibration & Generalization A 2-layer MLP head is calibrated with Monte Carlo dropout for reliable uncertainty estimates, and ALCoFM is pretrained on 15 U.S. cities with a classweighted loss, then fine-tuned on new regions us… view at source ↗

**Figure 2.** Road accident fatality rates (per 100,000 population) in various countries (2000–2022). (3) Spatio-Temporal Graph Neural Networks: GNNbased approaches model both spatial correlations and temporal dependencies in accident prediction. Early works such as DCRNN (Li et al., 2018) and STGCN (Yu et al., 2018) established the foundation for graph-based traffic forecasting. Subsequent studies introduced attention mechani… view at source ↗

**Figure 3.** presents some examples of map tiles retrieved from OSM at zoom level 14, which are used to approximate H3 zones (R = 7) in regions of Columbus, Ohio. These map images encapsulate critical environmental elements, including roads, intersections, buildings, and other infrastructure, offering valuable insights into potential accident-prone areas (Monsefi et al., 2022; Karimi Monsefi et al., 2023). By leveraging these ge… view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Pairwise correlation matrix of numerical and categorical features. 9 [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]

**Figure 5.** Figure 5: Distribution of accident vs. no-accident instances across different cities. 10 [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]

**Figure 6.** Figure 6: Cumulative explained variance from PCA across top 30 components. 11 [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]

**Figure 7.** Figure 7: XGBoost-based feature importance ranking. 12 [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]

**Figure 8.** Figure 8: t-SNE projection of the dataset in 2D space, colored by accident labels. 13 [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]

**Figure 9.** Figure 9: PCA projection of the dataset in 2D space, colored by accident labels. 14 [PITH_FULL_IMAGE:figures/full_fig_p026_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 20 canonical work pages

[5]

doi: 10.3390/s23229225

ISSN 1424-8220. doi: 10.3390/s23229225. URL https://www.mdpi.com/1424-8220/23/ 22/9225. Ding, C., Sun, S., and Zhao, J. Mst-gat: A mul- timodal spatial–temporal graph attention network for time series anomaly detection. Information Fusion, 89:527–536,

doi:10.3390/s23229225
[6]

doi: https://doi.org/10.1016/j.inffus.2022.08.011

ISSN 1566-2535. doi: https://doi.org/10.1016/j.inffus.2022.08.011. URL https://www.sciencedirect.com/ science/article/pii/S156625352200104X. Du, W., Cˆot´e, D., and Liu, Y . Saits: Self-attention-based imputation for time series. Expert Systems with Appli- cations, 219:119619, June

work page doi:10.1016/j.inffus.2022.08.011 2022
[7]

doi: 10.1016/j.eswa.2023.119619

ISSN 0957-4174. doi: 10.1016/j.eswa.2023.119619. URL http://dx.doi. org/10.1016/j.eswa.2023.119619. Fang, Z., Long, Q., Song, G., and Xie, K. Spatial- temporal graph ode networks for traffic flow forecast- ing. In Proceedings of the 27th ACM SIGKDD Con- ference on Knowledge Discovery & Data Min- ing, KDD ’21, pp. 364–373. ACM, August

arXiv 2023
[10]

URL https://doi.org/ 10.1023/A:1010933404324

9 ALCo-FM 1023/A:1010933404324. URL https://doi.org/ 10.1023/A:1010933404324. Chen, M., Yuan, H., Jiang, N., Bao, Z., and Wang, S. Ur- ban traffic accident risk prediction revisited: Regionality, proximity, similarity and sparsity. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM ’24, pp. 281–290. ACM, ...

arXiv
[12]

iihs.org/topics/fatality-statistics/ detail/yearly-snapshot

URL https://www. iihs.org/topics/fatality-statistics/ detail/yearly-snapshot. Latest data released by IIHS, accessed: 2025-01-28. Iowa State University. Iowa environmental mesonet (iem) asos-awos-metar data. https://mesonet.agron. iastate.edu/request/download.phtml. Ac- cessed: 2025-01-26. Jin, G., Li, F., Zhang, J., Wang, M., and Huang, J. Auto- mated di...

work page 2025
[13]

ISBN 9798400703621

Association for Comput- ing Machinery. ISBN 9798400703621. doi: 10.1145/ 3615900.3628769. URL https://doi.org/10. 1145/3615900.3628769. Khan, K., Zaidi, S. B., and Ali, A. Evaluating the nature of distractive driving factors towards road traffic accident. Civil Engineering Journal, 6(8):1555–1580,

work page arXiv
[14]

Kim, T.-w., Lee, H.-j., Jung, H.-J., Yang, J.-W., and Hong, E. J. Sst-gcn: The sequential based spatio-temporal graph convolutional networks for minute-level and road- level traffic accident risk predictio. arXiv preprint arXiv:2405.18602,

work page Pith review arXiv
[15]

Li, F., Feng, J., Yan, H., Jin, G., Yang, F., Sun, F., Jin, D., and Li, Y

doi: 10.1109/TITS.2024.3435995. Li, F., Feng, J., Yan, H., Jin, G., Yang, F., Sun, F., Jin, D., and Li, Y . Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution. ACM Transactions on Knowledge Discovery from Data, 17(1):1–21, 2023a. Li, L., Dou, Y ., and Zhou, J. Traffic accident detection based on multimodal knowled...

arXiv 2024

Show all 34 references

[16]

doi: 10.1609/aaai.v35i5. 16542. URL https://ojs.aaai.org/index. php/AAAI/article/view/16542. Li, Y ., Yu, R., Shahabi, C., and Liu, Y . Diffusion con- volutional recurrent neural network: Data-driven traffic forecasting,

doi
[19]

URL https://doi.org/ 10.1145/3557915.3560943

1145/3557915.3560943. URL https://doi.org/ 10.1145/3557915.3560943. Moosavi, S., Samavatian, M. H., Parthasarathy, S., and Ramnath, R. A countrywide traffic accident dataset.arXiv preprint arXiv:1906.05409,

1906
[20]

Autostg: Neural architecture search for predictions of spatio-temporal graph*

Pan, Z., Ke, S., Yang, X., Liang, Y ., Yu, Y ., Zhang, J., and Zheng, Y . Autostg: Neural architecture search for predictions of spatio-temporal graph*. In Pro- ceedings of the Web Conference 2021 , WWW ’21, pp. 1846–1855, New York, NY , USA,

2021
[21]

ISBN 9781450383127

Associa- tion for Computing Machinery. ISBN 9781450383127. doi: 10.1145/3442381.3449816. URL https://doi. org/10.1145/3442381.3449816. Song, C., Lin, Y ., Guo, S., and Wan, H. Spatial-temporal synchronous graph convolutional networks: A new frame- work for spatial-temporal net...
[22]

Sun, Y ., Zhu, D., Wang, Y ., Fu, Y ., and Tian, Z

URL https://arxiv.org/abs/1905.07799. Sun, Y ., Zhu, D., Wang, Y ., Fu, Y ., and Tian, Z. Gtc: Gnn-transformer co-contrastive learning for self- supervised heterogeneous graph representation. Neural Networks, 181:106645,

1905 arXiv
[24]

United States Postal Service

Accessed: 2025-01-13. United States Postal Service. Zip code lookup and address information. https://www.usps.com. Accessed: 2025-01-13. United States ZIP Codes. United states zip codes database. https://www.unitedstateszipcodes.org/. Accessed: 2025-01-13. U.S. Census Bureau. ...

2025 arXiv
[25]

Latest data released by WHO, accessed: 2025-01-28

URL https://apps.who.int/ gho/data/node.main.A997. Latest data released by WHO, accessed: 2025-01-28. Wu, K., Li, W., and Xiao, X. Accidentgpt: Large multi-modal foundation model for traffic accident analy- sis,

2025
[26]

Xia, C., Wang, X., Lv, F., Hao, X., and Shi, Y

URL https://arxiv.org/abs/1906.00121. Xia, C., Wang, X., Lv, F., Hao, X., and Shi, Y . Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ...

1906 arXiv
[27]

URL https://doi.org/ 10.24963/ijcai.2018/505

doi: 10.24963/ijcai.2018/505. URL https://doi.org/ 10.24963/ijcai.2018/505. Yu, L., Du, B., Hu, X., Sun, L., Han, L., and Lv, W. Deep spatio-temporal graph convolutional network for traf- fic accident prediction. Neurocomputing, 423:135–147,

2018 doi
[28]

Point of Interest (POI)

12 ALCo-FM Appendix A. Details of Dataset Descriptions A.1. Traffic Events Data To build a robust dataset for accident risk prediction, we leverage a comprehensive car accident dataset by Moosavi et al. (Moosavi et al., 2019), covering 49 states across the United States. This ...

2019
[29]

Type Description Amenity Denotes specific locations such as restaurants, libraries, colleges, bars, etc

Definition of Point-Of-Interest (POI) annotation tags based on OpenStreetMap (OSM). Type Description Amenity Denotes specific locations such as restaurants, libraries, colleges, bars, etc. Bump Represents speed bumps or humps designed to reduce vehicle speed. Crossing Indicate...

2016
[30]

This resolution results in hexagons with an approximate edge length of 2,604 meters and a total area of about 5.16 km 2, effectively covering the vicinity of the accident site. Once the hexagonal zoning is established, the center coordinates of each hexagonal region are used t...

2017
[31]

in regions of Columbus, Ohio. These map images encapsulate critical environmental elements, including roads, intersections, buildings, and other infrastructure, offering valuable insights into potential accident-prone areas (Monsefi et al., 2022; Karimi Monsefi et al., 2023). ...

2022
[32]

worst-case

Summary of temporal features derived from timestamps spanning 2016–2023. Numeric values used for categorical encoding are indicated in parentheses. Feature Description Season Categorized based on the month: Winter (0) [Dec, Jan, Feb]; Spring (1) [Mar, Apr, May]; Summer (2) [Ju...

2016
[33]

It addresses the slow query issue of classical kNN by using efficient nearest-neighbor lookups

: Scalable similarity search, well-suited for large datasets. It addresses the slow query issue of classical kNN by using efficient nearest-neighbor lookups. We chose k = 5after tuning on k ∈ {3, 5, 7, 9, 15}, as it yielded the lowest error. • SAITS (Du et al., 2023): A self-a...

2023
[34]

7 ALCo-FM Most missing values originate from demographic attributes for certain ZIP codes

offer an emerging alternative for tasks involving spatially grounded decisions, and could be explored in future expansions of our work. 7 ALCo-FM Most missing values originate from demographic attributes for certain ZIP codes. Since demographic data tends to be more correlated...

2001
[1993]

doi: https://doi.org/10.1016/0098-3004(93)90090-R

ISSN 0098-3004. doi: https://doi.org/10.1016/0098-3004(93)90090-R. URL https://www.sciencedirect.com/ science/article/pii/009830049390090R. Mohammadshirazi, A., Firoozsalari, A. N., Zhou, M., Kulshrestha, D., and Ramnath, R. Docparsenet: Ad- vanced semantic segmentation and oc...
[2016]

URL http: //dx.doi.org/10.1145/2939672.2939785

doi: 10.1145/2939672.2939785. URL http: //dx.doi.org/10.1145/2939672.2939785. Chen, Y ., Ren, K., Wang, Y ., Fang, Y ., Sun, W., and Li, D. Contiformer: Continuous-time transformer for irregular time series modeling. Advances in Neural Information Processing Systems, 36, 2024b...
[2017]

Guha Neogi, P

URL https://arxiv.org/abs/ 1603.08983. Guha Neogi, P. P. and Goswami, S. Force of gravity oriented classification technique in machine learning. In Sharma, N., Chakrabarti, A., Balas, V . E., and Martinovic, J. (eds.), Data Management, Analytics and Innovation, pp. 299– 310, S...

arXiv
[2018]

Ma´ckiewicz, A

URL https://arxiv.org/abs/ 1707.01926. Ma´ckiewicz, A. and Ratajczak, W. Principal compo- nents analysis (pca). Computers & Geosciences , 19(3):303–342,

arXiv
[2019]

Multi-scale cross- attention transformer encoder for event classification

Hammad, A., Moretti, S., and Nojiri, M. Multi-scale cross- attention transformer encoder for event classification. Journal of High Energy Physics, 2024(3):1–31,

2024
[2020]

Berhanu, Y ., Alemayehu, E., and Schr ¨oder, D. Examin- ing car accident prediction techniques and road traffic congestion: A comparative analysis of road safety and prevention of world challenges in low-income and high- income countries. Journal of advanced transportation, 20...

2023
[2021]

URL http://dx.doi

doi: 10.1145/3447548.3467430. URL http://dx.doi. org/10.1145/3447548.3467430. Gan, J., Yang, Q., Zhang, D., Li, L., Qu, X., and Ran, B. A novel voronoi-based spatio-temporal graph convolu- tional network for traffic crash prediction considering geographical spatial distributio...
[2023]

de Zarz `a, I., de Curt `o, J., Roig, G., and Calafate, C

URL https://arxiv.org/abs/2301.13338. de Zarz `a, I., de Curt `o, J., Roig, G., and Calafate, C. T. Llm multimodal traffic accident forecasting. Sensors, 23 (22),

arXiv
[2024]

doi: 10.1007/ s10115-024-02159-7

ISSN 0219-3116. doi: 10.1007/ s10115-024-02159-7. URL http://dx.doi.org/ 10.1007/s10115-024-02159-7 . Guo, S., Lin, Y ., Feng, N., Song, C., and Wan, H. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In Proceedings of the AAAI con- ...

doi
[2025]

doi: https://doi.org/10.1016/j.neunet.2024.106645

ISSN 0893-6080. doi: https://doi.org/10.1016/j.neunet.2024.106645. URL https://www.sciencedirect.com/ science/article/pii/S0893608024005690. Uber Technologies, I. Uber h3: Hexagonal hierarchi- cal geospatial indexing system. https://uber. github.io/h3/,

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [5]

doi: 10.3390/s23229225

ISSN 1424-8220. doi: 10.3390/s23229225. URL https://www.mdpi.com/1424-8220/23/ 22/9225. Ding, C., Sun, S., and Zhao, J. Mst-gat: A mul- timodal spatial–temporal graph attention network for time series anomaly detection. Information Fusion, 89:527–536,

doi:10.3390/s23229225

[2] [6]

doi: https://doi.org/10.1016/j.inffus.2022.08.011

ISSN 1566-2535. doi: https://doi.org/10.1016/j.inffus.2022.08.011. URL https://www.sciencedirect.com/ science/article/pii/S156625352200104X. Du, W., Cˆot´e, D., and Liu, Y . Saits: Self-attention-based imputation for time series. Expert Systems with Appli- cations, 219:119619, June

work page doi:10.1016/j.inffus.2022.08.011 2022

[3] [7]

doi: 10.1016/j.eswa.2023.119619

ISSN 0957-4174. doi: 10.1016/j.eswa.2023.119619. URL http://dx.doi. org/10.1016/j.eswa.2023.119619. Fang, Z., Long, Q., Song, G., and Xie, K. Spatial- temporal graph ode networks for traffic flow forecast- ing. In Proceedings of the 27th ACM SIGKDD Con- ference on Knowledge Discovery & Data Min- ing, KDD ’21, pp. 364–373. ACM, August

arXiv 2023

[4] [10]

URL https://doi.org/ 10.1023/A:1010933404324

9 ALCo-FM 1023/A:1010933404324. URL https://doi.org/ 10.1023/A:1010933404324. Chen, M., Yuan, H., Jiang, N., Bao, Z., and Wang, S. Ur- ban traffic accident risk prediction revisited: Regionality, proximity, similarity and sparsity. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM ’24, pp. 281–290. ACM, ...

arXiv

[5] [12]

iihs.org/topics/fatality-statistics/ detail/yearly-snapshot

URL https://www. iihs.org/topics/fatality-statistics/ detail/yearly-snapshot. Latest data released by IIHS, accessed: 2025-01-28. Iowa State University. Iowa environmental mesonet (iem) asos-awos-metar data. https://mesonet.agron. iastate.edu/request/download.phtml. Ac- cessed: 2025-01-26. Jin, G., Li, F., Zhang, J., Wang, M., and Huang, J. Auto- mated di...

work page 2025

[6] [13]

ISBN 9798400703621

Association for Comput- ing Machinery. ISBN 9798400703621. doi: 10.1145/ 3615900.3628769. URL https://doi.org/10. 1145/3615900.3628769. Khan, K., Zaidi, S. B., and Ali, A. Evaluating the nature of distractive driving factors towards road traffic accident. Civil Engineering Journal, 6(8):1555–1580,

work page arXiv

[7] [14]

Kim, T.-w., Lee, H.-j., Jung, H.-J., Yang, J.-W., and Hong, E. J. Sst-gcn: The sequential based spatio-temporal graph convolutional networks for minute-level and road- level traffic accident risk predictio. arXiv preprint arXiv:2405.18602,

work page Pith review arXiv

[8] [15]

Li, F., Feng, J., Yan, H., Jin, G., Yang, F., Sun, F., Jin, D., and Li, Y

doi: 10.1109/TITS.2024.3435995. Li, F., Feng, J., Yan, H., Jin, G., Yang, F., Sun, F., Jin, D., and Li, Y . Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution. ACM Transactions on Knowledge Discovery from Data, 17(1):1–21, 2023a. Li, L., Dou, Y ., and Zhou, J. Traffic accident detection based on multimodal knowled...

arXiv 2024