Pith. sign in

REVIEW 3 major objections 5 minor 26 references

MambaLSTM, a framework that couples a Mamba state-space block with a rebuilt LSTM-style gating unit and multi-scale patch embeddings, claims consistently higher accident-risk prediction F1 than previous models on three city datasets.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 17:06 UTC pith:7R7DBJKL

load-bearing objection Plausible new Mamba-LSTM hybrid for accident risk prediction, with real F1 gains on two datasets and a load-bearing inconsistency in the core cell equations that needs fixing before the empirical claim can be accepted. the 3 major comments →

arxiv 2607.18353 v1 pith:7R7DBJKL submitted 2026-07-20 cs.LG cs.AI

MambaLSTM: A Spatio-Temporal Framework for Enhanced Traffic Accident Risk Prediction

Zhen Yu , Yachao Yuan , Zixiang Peng , Muting Li , Thar Baker This is my paper
classification cs.LG cs.AI
keywords traffic accident risk predictionspatio-temporal modelingMambaLSTMstate-space modelssqueeze-and-excitation attentionmulti-scale patch embeddingurban computing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper proposes MambaLSTM, a framework for predicting traffic accident risk in city grids one time step ahead. It claims that by fusing temporal features in a squeeze-and-excitation attention step, converting spatial inputs into multi-scale patch embeddings, running a Mamba state-space block to capture global spatial correlations, and rebuilding the LSTM gating mechanism around that block, the model outperforms existing graph, Transformer, and Mamba baselines. On accident data from New York, Chicago, and a US-wide sample, the reported F1 scores are 59.52, 32.92, and 52.2 respectively, with the largest margin over the best baseline on NYC (about +2.7 points). The paper also reports that each module is necessary: removing any one lowers all evaluation metrics in the ablation study.

Core claim

The central claim is that state-space models and LSTM-style recurrent units can be combined into a single cell for grid-based traffic-risk forecasting, and that the resulting architecture is more accurate than models that treat spatial and temporal modeling separately. The specific design uses a squeeze-and-excitation temporal feature fusion module to rescale spatial features with channel attention driven by sinusoidal time embeddings, a multi-scale convolution patch embedding to capture local and neighborhood correlations, a Mamba block to encode global spatial semantics through a dual-branch linear-complexity state-space scan, and a MambaLSTM unit whose gates are produced by an MLP followi

What carries the argument

The load-bearing object is the MambaLSTM unit: a recurrent cell built on ConvLSTM but with the conventional weight matrices replaced by stacked Mamba blocks followed by an MLP. The concatenation of the current feature sequence and the previous hidden state passes through the Mamba block(s), whose output is split into four chunks; three chunks become sigmoid-activated input, forget, and output gates, and the fourth is tanh-activated as a candidate state. As written, the cell-state update forms the new contribution as the product of the input and output gate activations, while the candidate state is computed but does not enter the update. Supporting machinery includes the SE-TFF module, which

Load-bearing premise

The reported performance advantage assumes that the MambaLSTM unit described in the equations is exactly the model that was trained; since the printed cell-state update does not use the candidate state, the actual behavior of the unit is not fully determined by the paper alone.

What would settle it

Inspect the released implementation's cell-state update: if it follows Eq. (15) and ignores the candidate state, then the architecture is a gated residual cell, and retraining on the NYC benchmark with the reported hyperparameters will reveal whether the F1 of 59.52 can be reproduced or whether the text differs from the code.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The model is grid-based and does not require a hand-built road network graph, so the same framework can be applied to any city where mobility flows, accident records, and weather features are available.
  • With a prediction time of 1.76 seconds for an entire city on the Chicago benchmark, the architecture is fast enough for real-time urban operation despite being slower than a plain ConvLSTM.
  • The ablation study indicates that the attention-based temporal fusion and multi-scale patch embedding contribute to accuracy, not just the Mamba block, so the gains are not solely from switching to state-space models.
  • The reported margins are largest on NYC (+2.7 F1 points over the strongest baseline) and smaller on the US sample, suggesting the architecture's benefit varies with data density or accident sparsity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the cell-state update indeed omits the candidate state, the MambaLSTM unit is better described as a gated residual cell than as an LSTM; whether using the candidate state would improve or degrade the reported results is untested in the paper.
  • The same recipe—Mamba blocks inside a recurrent cell with channel-attention temporal fusion—could transfer to other spatio-temporal forecasting problems such as crime incident prediction or emergency call volume, where cyclical time effects and global spatial contagion matter.
  • Because the paper's text leaves grid dimensions, time-slot length, and split details unspecified, re-running the authors' code with the reported hyperparameters would settle whether the F1 values are stable across random seeds; the framework's robustness claim would be strengthened by such a check.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes MambaLSTM, a spatio-temporal deep-learning framework for grid-based traffic accident risk prediction. It combines a squeeze-and-excitation temporal feature fusion module (SE-TFF), a multi-scale convolutional patch embedding (PE-MSC), a dual-branch Mamba block for global spatial semantics, and a recurrent 'MambaLSTM' unit that reconstructs LSTM gating using Mamba blocks. Experiments on NYC, Chicago, and a US-Accidents sample compare F1-score, AUC-PR, and AUC-ROC against ten baselines, reporting the best F1 on all three datasets and best AUC-ROC on NYC, with ablations and an efficiency table. The code is released.

Significance. If the results hold, MambaLSTM would be a competitive new architecture for traffic accident risk prediction, combining state-space models with a recurrent gating mechanism and showing gains over strong baselines such as ST-GCN, MVMT-STN, and C-ViT. The paper has several strengths: it uses three real-world public datasets, compares against a broad set of recent baselines, includes an ablation study, reports parameter counts and inference times, and releases code. However, the significance is currently limited by reproducibility gaps and an internal inconsistency in the core recurrent unit, both of which need to be resolved before the empirical claims can be independently verified.

major comments (3)
  1. [§4.4, Eq. (15)] The core MambaLSTM unit is internally inconsistent as written. Eq. (12) chunks G into four gates, Eq. (14) computes the candidate state G̃c = tanh(Gc), but Eq. (15) updates C_t = G̃f ∘ C_{t-1} + G̃i ∘ G̃o, which omits G̃c and instead multiplies the input gate by the output gate. This contradicts the prose description in the same section, which states that the input gate and candidate state jointly update the cell. If the released code matches Eq. (15), then the architecture is not an LSTM-like unit and the prose is wrong; if the code matches the prose, the equations are wrong. Either way, independent verification of the central novel component is impossible from the manuscript. Please correct the equations or the text, and confirm which update is actually implemented.
  2. [§5.1 / §3] The experimental setup is severely under-specified. The reader is not told the grid dimensions H and W, the length of the time slot, the number of historical steps Tk and Tr, the weather feature dimension d, the train/validation/test split, or the prediction horizon for any dataset. The per-dataset weighting coefficient β is reported only for NYC (1.25) and CHI (1.5), with no value for the USA dataset and no description of how β was selected beyond 'extensive experiments.' These details are essential for reproducing Table 1 and for ruling out selection bias in the reported metrics. Please provide a complete data-preprocessing and training-protocol table or appendix.
  3. [§5.2, Table 1] The abstract and §5.2 claim that MambaLSTM 'consistently outperforms other models in nearly all metrics.' This is not supported on the USA dataset: the proposed model has AUC-PR 56.77, below MVMT-STN (57.29) and GSNet (56.81), and AUC-ROC 91.72, below ST-GCN (92.17). The overclaim should be tempered, and the authors should either report statistical significance (e.g., multiple seeds with means and standard deviations) or clearly identify the metrics for which the improvement is not consistent across all three datasets.
minor comments (5)
  1. [§4.1] The module is named SE-TFF in the title and Fig. 2, but the text refers to it as 'SE-FEF' in the first paragraph. Please unify the nomenclature.
  2. [§5.2, Ablation Study] The ablation results are only shown in Fig. 7 without numeric values, dataset names, or error bars. Since M3 (replacing MambaLSTM with ConvLSTM) is the only direct evidence for the novel recurrent unit, please provide exact numbers and state which two datasets are used.
  3. [§5.2, Efficiency Comparison] Table 2 reports parameter counts and prediction times on CHI only. It would be useful to clarify the hardware and batch size used, and whether the prediction time includes data loading and preprocessing.
  4. [§3, Eq. (2)] The severity encoding R2 = 1 + P_injured + 2*P_dead is introduced without justification or citation. Since this is a label construction choice, please discuss its validity and sensitivity, even briefly.
  5. [References] Reference [17] is dated 2026; please verify the publication status and year. Also, reference [2] is a route-finding paper, not a traffic accident prediction baseline; please check whether it is the intended citation for HA.

Circularity Check

0 steps flagged

No significant circularity: the model is evaluated on public datasets against external baselines; no load-bearing self-citation or fitted quantity is renamed as a prediction.

full rationale

The paper's derivation chain is not circular. MambaLSTM is defined by explicit equations in §4 (SE-TFF, PE-MSC, Mamba block, MambaLSTM unit), and its performance claim is tested by next-step prediction on NYC, CHI, and US-Accidents against independent baselines (HA, ARIMA, XGBoost, LSTM, ConvLSTM, GSNet, ST-GCN, MVMT-STN, MG-TAR, C-ViT, ST-TAR, Mamba). No reference is to the current authors' prior work, so there is no self-citation chain, uniqueness theorem, or ansatz-smuggling-via-citation. The only fitted quantities are standard training hyperparameters (beta in the weighted MSE loss, channel compression ratio r, model widths); beta is described as set by "extensive experiments" but it is a loss weight, not a target prediction, so reporting model performance after optimizing with it is not a reduction of the result to an input. The central equation inconsistency — Eq. (14) computes G̃c but Eq. (15) updates C_t without it — is a reproducibility/correctness flaw in the manuscript, not a circular derivation; it does not make the evaluation circular. Accordingly the circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The ledger shows this is an empirical architecture paper: there is no free derivation, so the contribution is the design plus benchmark results. The central claim depends on standard supervised-learning domain assumptions (predictability of accident risk from the chosen input tensors, taxi mobility as a proxy) and on a set of unstated hyperparameters and data-pipeline choices (grid geometry, time-slot length, history split Tk/Tr, weather dimension d, per-dataset loss weight β) that are fitted or chosen by hand. The severity encoding (Eq. 2) and the flattened scan order for the SSM are ad hoc choices without external justification. No physical entities (particles, forces) are postulated; the only invented objects are architectural components, each carrying falsifiable handles via the released code and held-out benchmarks.

free parameters (6)
  • Weighted-MSE coefficient β = 1.25 (NYC), 1.5 (CHI); USA not reported
    Set "by extensive experiments" (§5.1) per dataset; changes the training objective and hence all reported metrics.
  • SE-TFF channel compression ratio r = not reported
    Eq. (4) bottleneck-MLP reduction ratio; standard SE knob, value unstated.
  • Stacked Mamba block count (MMB depth) = not reported
    "MMB represents multiple mamba blocks" (§4.4); depth determines global-context modeling capacity and is unstated.
  • Hidden/latent dimensionality C and MLP widths = not reported
    Used throughout Eqs. (4)–(16) (LP(C,C/2), LP(C,C), etc.); unstated in §5.
  • Grid geometry H×W, time-slot length, Tk/Tr, weather dimension d = not reported
    Defined in §3 but never instantiated; the entire tensor pipeline and the claim's data dependence rest on these values.
  • USA dataset subset (500K sample, 2022 extraction) = 113K accidents
    Subsampling choice for the US benchmark with no justification; can materially alter accident-risk statistics and all USA rows.
axioms (6)
  • domain assumption Next-step accident risk is predictable from the previous T steps of mobility, accident history, weather, and calendar features.
    §3 problem formulation; the core supervised-learning premise, standard in the field, treated as given.
  • domain assumption Taxi trajectories are a valid proxy for human mobility relevant to accident risk.
    §3: "taxi trajectory data is utilized as a proxy for human mobility"; no evidence offered that taxi flow tracks the population movement that generates accidents.
  • domain assumption An H×W grid discretization and fixed time slots preserve accident-risk structure.
    §3: city divided into c_{i,j} cells; no sensitivity analysis, and H/W/time-slot values are never stated.
  • ad hoc to paper Severity encoding R2 = 1 + P_injured + 2·P_dead is a valid linear risk magnitude.
    §3, Eq. (2): arbitrary weights (injury 1, fatality 2) with no external calibration.
  • ad hoc to paper Serializing 2D spatial features into a 1D sequence for the SSM is order-meaningful.
    §4.2/§4.3: flattened spatial features are fed to the SSM without specifying scan order over the grid; SSM output depends on sequence order, which is unexamined.
  • domain assumption Weighted MSE with per-dataset β is an appropriate training objective for imbalanced accident data.
    §5.1: chosen for highly imbalanced risk data; β is fitted per dataset with no stated criterion such as metric-targeted selection.
invented entities (3)
  • MambaLSTM unit independent evidence
    purpose: Recurrent cell modeling long- and short-term temporal dependencies while Mamba blocks inside the gates supply global spatial context (Eqs. 11–16).
    New architectural unit with falsifiable handles (released code, reproducible benchmark numbers), but the Eq. (14)/(15) inconsistency means the implemented unit may not match the described one.
  • Dual-branch Mamba block variant independent evidence
    purpose: Captures global spatial semantics by fusing an SSM branch and a 1×1-conv branch (§4.3).
    Modification of the Mamba block from ref [26]; reproducible via code; independent evidence is the released code and the ablation/benchmark results.
  • SE-TFF and PE-MSC modules independent evidence
    purpose: Temporal-aware feature fusion (SE-TFF) and local multi-scale patch embedding (PE-MSC).
    Compositions of well-known blocks (SE-Net [19], multi-scale convolution); listed for completeness; evidence is the ablation study showing each module's contribution.

pith-pipeline@v1.3.0-alltime-deepseek · 9080 in / 25604 out tokens · 259807 ms · 2026-08-01T17:06:30.682843+00:00 · methodology

0 comments
read the original abstract

In traffic accident risk prediction, most studies overlook the extra noise that could be incorporated when fusing temporal features into spatial features, and some models struggle to capture global correlations among spatial regions. To address these challenges, we propose a novel traffic accident risk prediction framework named MambaLSTM. First, we develop a squeeze-and-excitation temporal feature fusion module to integrate temporal information without compromising spatio-temporal integrity. Second, we introduce a new patch embedding module for effectively capturing semantic relationships among spatially adjacent regions. Additionally, we introduce a Mamba block based on state-space models to model global spatial semantics in urban regions. Finally, we propose a MambaLSTM unit to efficiently capture long- and short-term temporal dependencies for identifying dynamic risk patterns. Extensive experiments on real-world datasets demonstrate the proposed model's superiority over state-of-the-art methods. The code is released at https://github.com/Zhenzovo/MambaLSTM.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 3 canonical work pages

  1. [1]

    IEEE Transactions on Intelligent Transportation Systems 19(3), 699 -- 710 (2018)

    Xiong, X., Chen, L., Liang, J.: A new framework of vehicle collision prediction by combin- ing svm and hmm. IEEE Transactions on Intelligent Transportation Systems 19(3), 699 -- 710 (2018). https://doi.org/10.1109/TITS.2017.2699191

    arXiv 2018

  2. [2]

    In: 22nd International Conference on Data Engineering (ICDE'06)

    Kanoulas, E., Du, Y., Xia, T., Zhang, D.: Finding fastest paths on a road network with speed patterns. In: 22nd International Conference on Data Engineering (ICDE'06). pp. 10 --10 (2006). https://doi.org/10.1109/ICDE.2006.71

    work page doi:10.1109/icde.2006.71 2006

  3. [3]

    Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks (2017), https://arxiv.org/abs/1609.02907

    Pith/arXiv arXiv 2017

  4. [4]

    In: Proceedings of the 27th International Joint Conference on Artificial Intelligence

    Yu, B., Yin, H., Zhu, Z.: Spatio -temporal graph convolutional networks: a deep learning framework for traffic forecasting. In: Proceedings of the 27th International Joint Conference on Artificial Intelligence. p. 3634--3640. IJCAI'18, AAAI Press (2018)

    2018

  5. [5]

    Advances in neural information processing systems 28 (2015)

    Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c.: Convolutional lstm network: A machine learning approach for precipitation nowcasting. Advances in neural information processing systems 28 (2015)

    2015

  6. [6]

    In: 2022 IEEE 25th International Conference on Intelligent Transpor- tation Systems (ITSC)

    Saleh, K., Grigorev, A., Mihaita, A.S.: Traffic accident risk forecasting using contextual vision transformers. In: 2022 IEEE 25th International Conference on Intelligent Transpor- tation Systems (ITSC). pp. 2086 --2092 (2022). https://doi.org/10.1109/ITSC55140.2022.9921978

    arXiv 2022

  7. [7]

    Knowledge-Based Systems 286, 111406 (2024)

    Alhaek, F., Liang, W., Rajeh, T.M., Javed, M.H., Li, T.: Learning spatial patterns and tem- poral dependencies for traffic accident severity prediction: A deep learning approach. Knowledge-Based Systems 286, 111406 (2024). 12 Z. Yu et al. https://doi.org/10.1016/j.knosys.2024.111406, https://www.sciencedirect.com/science/arti- cle/pii/S0950705124000418

    arXiv 2024

  8. [8]

    Information Sciences 697, 121767 (2025)

    Alhaek, F., Li, T., Rajeh, T.M., Javed, M.H., Liang, W.: Encoding global semantic and lo- calized geographic spatial-temporal relations for traffic accident risk prediction. Information Sciences 697, 121767 (2025). https://doi.org/10.1016/j.ins.2024.121767, https://www.sci- encedirect.com/science/article/pii/S0020025524016815

    arXiv 2025

  9. [9]

    In: Proceedings of Fifth International Conference on Soft Computing for Problem Solving: SocProS 2015, Volume 2

    Sharma, B., Katiyar, V.K., Kumar, K.: Traffic accident prediction model using support vec- tor machines with gaussian kernel. In: Proceedings of Fifth International Conference on Soft Computing for Problem Solving: SocProS 2015, Volume 2. pp. 1--10. Springer (2016)

    2015

  10. [10]

    Accident Analysis & Prevention 39(4), 657--670 (2007)

    Caliendo, C., Guida, M., Parisi, A.: A crash-prediction model for multilane roads. Accident Analysis & Prevention 39(4), 657--670 (2007)

    2007

  11. [11]

    In: 2009 international conference on measuring technology and mechatronics automation

    Lv, Y., Tang, S., Zhao, H.: Real -time highway traffic accident prediction based on the k - nearest neighbor method. In: 2009 international conference on measuring technology and mechatronics automation. vol. 3, pp. 547--550. IEEE (2009)

    2009

  12. [12]

    China Communications 19(2), 214 --224 (2022)

    Zhao, H., Li, X., Cheng, H., Zhang, J., Wang, Q., Zhu, H.: Deep learning -based prediction of traffic accidents risk for internet of vehicles. China Communications 19(2), 214 --224 (2022)

    2022

  13. [13]

    IEEE Transactions on Intelligent Transportation Systems 24(4), 3779--3794 (2023)

    Trirat, P., Yoon, S., Lee, J.G.: Mg-tar: Multi-view graph convolutional networks for traffic accident risk prediction. IEEE Transactions on Intelligent Transportation Systems 24(4), 3779--3794 (2023)

    2023

  14. [14]

    Advances in neural information processing systems 30 (2017)

    Wang, Y., Long, M., Wang, J., Gao, Z., Yu, P.S.: Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms. Advances in neural information processing systems 30 (2017)

    2017

  15. [15]

    arXiv preprint arXiv:2001.02908 (2020)

    Xu, M., Dai, W., Liu, C., Gao, X., Lin, W., Qi, G.J., Xiong, H.: Spatial-temporal transformer networks for traffic flow forecasting. arXiv preprint arXiv:2001.02908 (2020)

    Pith/arXiv arXiv 2001

  16. [16]

    Advances in neural information processing sys- tems 34, 22419--22430 (2021)

    Wu, H., Xu, J., Wang, J., Long, M.: Autoformer: Decomposition transformers with auto - correlation for long-term series forecasting. Advances in neural information processing sys- tems 34, 22419--22430 (2021)

    2021

  17. [17]

    arXiv preprint arXiv:2604.16859 (2026)

    He, D., Gao, Y., Jiang, B., Yan, H.: Gamma -net: Adaptive long-horizon traffic spatio-tem- poral forecasting model based on interleaved graph attention and multi -axis mamba. arXiv preprint arXiv:2604.16859 (2026)

    Pith/arXiv arXiv 2026

  18. [18]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Jin, X., Su, H., Liu, K., Ma, C., Wu, W., Hui, F., Yan, J.: Unimamba: Unified spatial-channel representation learning with group-efficient mamba for lidar-based 3d object detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 1407--1417 (2025)

    2025

  19. [19]

    In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 7132 --7141 (2018). https://doi.org/10.1109/CVPR.2018.00745

    arXiv 2018

  20. [20]

    In: 2008 International Symposium on Information Technology

    Zare Moayedi, H., Masnadi-Shirazi, M.A.: Arima model for network traffic prediction and anomaly detection. In: 2008 International Symposium on Information Technology. vol. 4, pp. 1--6 (2008). https://doi.org/10.1109/ITSIM.2008.4631947

    arXiv 2008

  21. [21]

    In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    Chen, T., Guestrin, C.: XGBoost: A scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 785 --794. KDD '16, Association for Computing Machinery, New York, NY, USA (2016). https://doi.org/10.1145/2939672.2939785

    arXiv 2016

  22. [22]

    Neural Networks 25, 70 --83 (2012)

    Monner, D., Reggia, J.A.: A generalized LSTM -like training algorithm for second -order recurrent neural networks. Neural Networks 25, 70 --83 (2012). https://doi.org/10.1016/j.neunet.2011.07.003, https://www.sciencedirect.com/science/arti- cle/pii/S0893608011002036 MambaLSTM: A Spatio-Temporal Framework 13

    work page doi:10.1016/j.neunet.2011.07.003 2012

  23. [23]

    In: 2021 AAAI Con- ference on Artificial Intelligence (AAAI'21) (2021)

    Wang, B., Lin, Y., Guo, S., Wan, H.: GSNet: Learning spatial -temporal correlations from geographical and semantic aspects for traffic accident risk forecasting. In: 2021 AAAI Con- ference on Artificial Intelligence (AAAI'21) (2021)

    2021

  24. [24]

    IEEE Transactions on Knowledge and Data Engineer- ing 35(12), 12323--12336 (2023)

    Wang, S., Zhang, J., Li, J., Miao, H., Cao, J.: Traffic accident risk prediction via multi-view multi-task spatio-temporal networks. IEEE Transactions on Knowledge and Data Engineer- ing 35(12), 12323--12336 (2023). https://doi.org/10.1109/TKDE.2021.3135621

    arXiv 2023

  25. [25]

    In: Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25

    Wang, H., Chen, L., Shang, S., Han, P., Jensen, C.S.: ST -TAR: An Efficient Spatio -Tem- poral Learning Framework for Traffic Accident Risk Forecasting. In: Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25. pp. 7778- -7785. International Joint Conferences on Artificial Intelligence Organization (2025). h...

    work page doi:10.24963/ijcai.2025/865 2025

  26. [26]

    arXiv preprint arXiv:2312.00752 (2024), https://arxiv.org/abs/2312.00752

    Gu, A., Dao, T.: Mamba: Linear -Time Sequence Modeling with Selective State Spaces. arXiv preprint arXiv:2312.00752 (2024), https://arxiv.org/abs/2312.00752

    Pith/arXiv arXiv 2024