Pith. sign in

REVIEW 3 major objections 3 minor 1 cited by

Neural Spatiotemporal Point Processes: Trends and Challenges

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This review claims to be the first comprehensive survey of neural spatiotemporal point processes, unifying existing work around a shared autoregressive likelihood framework and cataloging design choices, applications, and open challenges.

desk verdict A useful survey of neural STPPs with a sensible taxonomy, but the 'first comprehensive review' claim rests on an undocumented literature search and a noticeable self-citation pattern. read the letter →

arxiv 2502.09341 v1 pith:SLZELMGQ submitted 2025-02-13 cs.LG cs.AI

Sumantrak Mukherjee , Mouad Elhamdi , George Mohler , David A. Selby , Yao Xie , Sebastian Vollmer , Gerrit Grossmann This is my paper · ORCID
classification cs.LGcs.AI
keywords neuralspatiotemporalpointprocessesconditionalintensityfunctiondeeplearningsurveyeventpredictionHawkeskernelmethodsdiffusionmodelsuncertaintyquantification
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

The paper sets out to establish that neural spatiotemporal point processes (STPPs) have matured into a coherent field that can be reviewed as a whole, and that no prior survey has done this comprehensively. It argues that deep learning overcomes the limits of traditional parametric STPP models by encoding event histories into latent states, learning non-stationary kernels, and avoiding intractable likelihood integrals. The review organizes the literature by design choices—history encoders, kernels, neural architectures, training objectives, and evaluation metrics—and shows that most methods reduce to a common autoregressive likelihood with a neural conditional intensity function. If right, researchers get a structured map of the field and a prioritized list of open problems, with missing benchmarks and reproducibility the most pressing. The load-bearing premise is that the literature search was comprehensive and representative.

What carries the argument

The central object is the conditional intensity function $\lambda^*(t, s \mid H_t)$, defined as the limiting rate of events in a small ball $B(s, \Delta s) \times [t, t+\Delta t)$, and the autoregressive likelihood that factorizes over observed events and the probability of no events after the last one. The paper shows that most neural STPPs instantiate this via a history encoder (RNN/LSTM or Transformer) producing a latent state, and a kernel-based intensity $\lambda^* = \mu + \sum_{(t',s') \in H_t} K(t', t, s', s)$, where $K$ is often decomposed into neural basis functions over time and space. This machinery lets the field be compared at the level of design choices rather than isolated papers.

What would settle it

Locating a peer-reviewed survey published before February 2025 that already covers neural spatiotemporal point processes with comparable breadth would falsify the novelty claim; likewise, a systematic, reproducible literature search that surfaces a substantial body of neural STPP papers absent from this review would undermine the comprehensiveness claim.

Watch

Extended reading notes

Core claim

The central claim is that neural STPPs can be unified under a single framework: an autoregressive likelihood over event sequences $f(X) = \prod_{i=1}^n f_{\text{pred}}(t_i, s_i \mid H_{t_i}) \cdot (1 - F_{\text{pred}}(T \mid H_{t_n}))$, where a neural network parameterizes the predictive distribution or its conditional intensity function $\lambda^*(t, s \mid H_t)$. Within this frame, the paper distinguishes methods by how they encode spatial structure (raw coordinates, learned embeddings, graphs, non-Euclidean spaces), how they specify the influence of past events (parametric kernels, neural basis decompositions $K = \sum_{r,l} \alpha_{rl} \psi_l \phi_r$, mixture models, diffusion), and how they train and evaluate. It also claims that multi-event prediction, score-matching and automatic integration, and uncertainty quantification are the emerging frontiers, while reproducibility and standardized benchmarks are the main barriers.

Load-bearing premise

The survey is only as comprehensive as its literature search, which used keyword queries, citation tracking, and known temporal-point-process works without a documented protocol for databases, dates, or inclusion criteria.

Editorial extensions

If this is right

  • Researchers can place any new neural STPP model in the taxonomy by identifying its history encoder, kernel parameterization, and training objective, which makes apples-to-apples comparisons easier.
  • The field's next bottleneck is not model capacity but shared infrastructure: the review treats missing standardized datasets and benchmark libraries as a major barrier to progress.
  • Methods that avoid likelihood integration—score matching, automatic integration, and diffusion-based sequence generation—are positioned as the most promising routes to scalable training.
  • Applications in crime, traffic, epidemiology, and natural disasters are mature enough that the next push must address interpretability, causality, and uncertainty before real deployment.

Reading between the lines

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

  • My inference: the taxonomy predicts convergence between kernel-based interpretable models and attention-based flexible models, since both appear as complementary ways to parameterize the same influence kernel.
  • My inference: the lack of a unified benchmark is likely to be filled by a community dataset effort mirroring the one that standardized purely temporal point processes, and the paper's design-choice frame gives that effort a ready checklist.
  • My inference: techniques currently proven in purely temporal point processes—such as flow-based and diffusion-based generative training—will transfer to STPPs faster once spatial encoding is decoupled from temporal encoding, which the paper notes is often treated independently.
  • My inference: if score-based pseudolikelihood estimation matures, it could make neural STPPs practical for safety-critical settings by providing confidence regions, a capability the paper flags as largely missing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. This manuscript is a survey of neural spatiotemporal point processes (STPPs). It introduces the standard probabilistic formulation of STPPs, reviews neural history encoders, single- and multi-event prediction models, parameter estimation and inference methods, evaluation metrics, and application domains such as crime, traffic, epidemiology, and natural disasters. It closes with a list of open challenges and an ethical statement. The paper's stated contribution is to be the first comprehensive review of neural STPPs, unifying design choices and identifying gaps in the literature.

Significance. If the survey's coverage is indeed comprehensive and representative, it would provide a useful structured map of a rapidly growing field. The paper has several strengths: the mathematical background on intensities and likelihoods is standard and clearly presented; Table 1 provides a helpful consolidated summary of evaluation metrics; the discussion of inference beyond maximum likelihood (score matching, automatic integration, imitation learning) is informative; and the inclusion of an ethical statement is commendable. The survey also correctly identifies reproducibility and benchmarking as important open problems. However, the central claim of comprehensiveness and novelty is not currently verifiable from the manuscript because the literature search methodology is not documented. The contribution is primarily organizational, so the validity of the survey rests on the completeness and representativeness of its coverage.

major comments (3)
  1. [1 (Introduction, 'Scope and structure')] The claim that 'no prior survey has comprehensively examined these aspects in this context' is load-bearing, but the search methodology is described in only one sentence: 'keyword-based queries, citation tracking, and seminal works in neural temporal point processes.' There is no list of databases, no query strings, no date range, no inclusion or exclusion criteria, and no screening or selection flowchart. Without a reproducible protocol, a reader cannot verify the comprehensiveness of the review or assess whether relevant 2023-2025 work or prior surveys were missed. Please add a detailed methodology subsection, and consider softening the novelty claim to 'to our knowledge' with an explicit comparison to the cited prior reviews by Bernabeu et al. [2024] and Wikle and Zammit-Mangion [2023] explaining why they are not considered comprehensive.
  2. [1 (Introduction) and throughout] The reviewed literature appears to concentrate heavily on a small set of research groups (e.g., Dong, Xie, Zhu, Okawa, Yuan, Li, and immediate collaborators), and many of the sections are organized around those works. This raises a selection-bias risk: the taxonomy and the list of open challenges may over-represent one intellectual cluster rather than the full field. For a survey whose main contribution is comprehensiveness, please provide a coverage table or a per-category enumeration of works, and state explicitly how applications and methods were chosen. If the concentration reflects the actual literature, that should be demonstrated rather than assumed.
  3. [3.2 (Single Event Prediction, Eq. (3))] There is a subscript inconsistency between Eq. (3) and its explanatory text. Equation (3) writes the kernel as K(t', t, s', s) = sum_l phi^{(l)}_{s'} g(t, t', s, s' | Sigma^{(l)}_{s'}, mu^{(l)}_{s'}), but the text states that the network embeds spatial coordinates s to generate location-specific parameters mu^{(l)}_s, Sigma^{(l)}_s, and phi^{(l)}_s. Please clarify whether the location-specific parameters are indexed by the source location s' or the target location s, and make the notation consistent throughout the subsection.
minor comments (3)
  1. [5 (Evaluation Metrics)] The sentence 'Prediction accuracy (ACC) is useful for event count estimation but, unlike metrics considering location and time, it assesses the accuracy of event counts only.' is immediately followed by a near-duplicate sentence: 'Prediction accuracy (ACC) is useful for event count estimation but doesn't account for spatial and temporal precision.' Please remove the duplicate.
  2. [3.2 (Single Event Prediction, discussion of Li et al. [2024])] The sentence '...while also providing uncertainty estimates the score function represents the gradient of the logarithm of the conditional spatial distribution.' is missing punctuation and reads as a run-on. Please split it into two sentences and clarify the relationship between the uncertainty estimates and the score function.
  3. [7 (Open Challenges)] The discussion of reproducibility mentions 'unified libraries like Xue et al. [2024]' but does not give the library name (EasyTPP) in the text. Adding the name and a brief description would make the point concrete for readers who are not familiar with that work.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a literature review, and its central claims do not reduce to fitted inputs or self-citations.

full rationale

This manuscript is a survey, not a derivation. The equations it presents (e.g., the likelihood factorization, the intensity function definition, and the kernel forms) are standard definitions or representative model forms taken from the cited literature; they are not derived from, nor do they define, the survey's conclusions. The central claim—'To our knowledge, no prior survey has comprehensively examined these aspects in this context'—is a factual assertion about the state of the literature, not a consequence of any equation or fitted parameter. The authors do cite their own prior work (e.g., Dong, Xie, Zhu, Mohler) as examples of neural STPP methods, but these citations are illustrative of existing approaches rather than load-bearing premises that force a prediction to follow by construction. The main weakness is that the literature search is documented only as 'keyword-based queries, citation tracking, and seminal works in neural temporal point processes,' without a reproducibility protocol; this affects the verifiability of the 'first comprehensive survey' claim and may indicate selection bias, but it is a completeness and correctness risk, not a circularity. No step in the paper reduces to its own inputs by definition, no fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported solely through self-citation. Therefore the circularity score is 0.

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

As a review, the paper introduces no free parameters or invented entities. It rests on the correctness of the cited literature, the validity of standard point process theory, and the authors' judgments about taxonomy completeness and research gaps.

assumptions (4)
  • domain assumption The reviewed papers' methods and results are accurately represented by the survey's summaries and equations.
    The survey's value depends on the fidelity of its summaries, but the authors provide no reproducibility artifacts to verify each summary.
  • domain assumption The taxonomy of spatial encodings (Euclidean, graph-based, non-Euclidean) is a meaningful and exhaustive partition of existing neural STPP methods.
    The survey groups papers by spatial encoding, but the Figure 2 caption acknowledges these categorizations are not strictly defined.
  • domain assumption The identified open challenges (reproducibility, benchmarking, architectures, applicability, causality and uncertainty) are the most pressing gaps in the field.
    This prioritization is the authors' judgment, not derived from a quantitative analysis.
  • standard math The conditional intensity function and the autoregressive likelihood decomposition in Section 2 are valid for spatiotemporal point processes.
    This is established point process theory from Daley and Vere-Jones, which the survey uses as its mathematical basis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Spatiotemporal Point Processes: Trends and Challenges." pith.science (2026). https://pith.science/paper/SLZELMGQ

@misc{pith2026250209341,
  author       = {Pith},
  title        = {Pith review of: Neural Spatiotemporal Point Processes: Trends and Challenges},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SLZELMGQ}},
  note         = {Machine review of arXiv:2502.09341}
}
read the original abstract

Spatiotemporal point processes (STPPs) are probabilistic models for events occurring in continuous space and time. Real-world event data often exhibit intricate dependencies and heterogeneous dynamics. By incorporating modern deep learning techniques, STPPs can model these complexities more effectively than traditional approaches. Consequently, the fusion of neural methods with STPPs has become an active and rapidly evolving research area. In this review, we categorize existing approaches, unify key design choices, and explain the challenges of working with this data modality. We further highlight emerging trends and diverse application domains. Finally, we identify open challenges and gaps in the literature.

Figures

Figures reproduced from arXiv: 2502.09341 by the authors.

Figure 1
Figure 1. Schematic of the autoregressive construction of an STPP [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. A timeline of the reviewed methodological and application-focused works, along with an overview of neural architectures and [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Manifold Constrained Conformal Prediction for Spatial Events

    stat.ML 2026-07 conditional novelty 6.0 of 10

    Manifold-constrained conformal prediction with sliced Wasserstein scores yields near-nominal coverage and lower energy/manifold distances for spatial event clouds than HDR or generative baselines.

Reference graph

Works this paper leans on

12 extracted references · 6 canonical work pages · cited by 1 Pith paper

  1. [2024]

    Spatio-temporal point processes with deep non-stationary kernels

    Zheng Dong, Xiuyuan Cheng, and Yao Xie. Spatio-temporal point processes with deep non-stationary kernels. arXiv preprint arXiv:2211.11179,

    work page Pith review arXiv

  2. [2023]

    Spatio-temporal- network point processes for modeling crime events with landmarks

    Zheng Dong, Jorge Mateu, and Yao Xie. Spatio-temporal- network point processes for modeling crime events with landmarks. arXiv preprint arXiv:2409.10882,

    arXiv

  3. [5]

    Quantifying the vanishing gradient and long distance dependency problem in recursive neural networks and recursive LSTMs

    Phong Le and Willem Zuidema. Quantifying the vanishing gradient and long distance dependency problem in recur- sive neural networks and recursive lstms. arXiv preprint arXiv:1603.00423,

    work page Pith review arXiv

  4. [6]

    Understanding the spread of covid-19 epidemic: A spatio-temporal point process view

    Shuang Li, Lu Wang, Xinyun Chen, Yixiang Fang, and Yan Song. Understanding the spread of covid-19 epidemic: A spatio-temporal point process view. arXiv preprint arXiv:2106.13097,

    arXiv

  5. [7]

    An empirical study: extensive deep temporal point process

    Haitao Lin, Cheng Tan, Lirong Wu, Zhangyang Gao, Stan Li, et al. An empirical study: extensive deep temporal point process. arXiv preprint arXiv:2110.09823,

    arXiv

  6. [2003]

    Atlanta Gun Violence Modeling via Nonstationary Spatio-temporal Point Processes

    Zheng Dong and Yao Xie. Atlanta gun violence modeling via nonstationary spatio-temporal point processes. arXiv preprint arXiv:2408.09258,

    work page Pith review arXiv

  7. [2016]

    Beyond Hawkes: Neural Multi-event Forecasting on Spatio-temporal Point Processes

    Negar Erfanian, Santiago Segarra, and Maarten de Hoop. Beyond hawkes: Neural multi-event forecasting on spatio-temporal point processes. arXiv preprint arXiv:2211.02922,

    work page Pith review arXiv

  8. [2017]

    Discovering latent structural causal models from spatio-temporal data

    Kun Wang, Sumanth Varambally, Duncan Watson-Parris, Yi-An Ma, and Rose Yu. Discovering latent structural causal models from spatio-temporal data. arXiv preprint arXiv:2411.05331,

    arXiv

Show all 12 references
  1. [2019]

    Lecture notes: Temporal point processes and the conditional intensity function

    Jakob Gulddahl Rasmussen. Lecture notes: Temporal point processes and the conditional intensity function. arXiv preprint arXiv:1806.00221,

    arXiv

  2. [2020]

    Neural temporal point pro- cesses: A review

    Oleksandr Shchur, Ali Caner T ¨urkmen, Tim Januschowski, and Stephan G ¨unnemann. Neural temporal point pro- cesses: A review. arXiv preprint arXiv:2104.03528,

    arXiv

  3. [2021]

    Exploring generative neural temporal point process

    Haitao Lin, Lirong Wu, Guojiang Zhao, Pai Liu, and Stan Z Li. Exploring generative neural temporal point process. arXiv preprint arXiv:2208.01874,

    arXiv

  4. [2022]

    Transformer hawkes process

    Simiao Zuo, Haoming Jiang, Zichong Li, Tuo Zhao, and Hongyuan Zha. Transformer hawkes process. In Inter- national conference on machine learning , pages 11692– 11702. PMLR, 2020

    2020

Pith tools

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