{"id":"ec75ae9b-a563-4beb-8556-e1a1333c2fcf","arxiv_id":"2502.09341","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured survey of neural spatiotemporal point processes that unifies design choices, reviews applications, and identifies open research challenges.","lead":"This paper reviews computer models that predict events at specific places and times, such as crimes, earthquakes, and disease outbreaks. It organizes existing approaches into categories, compares their design choices, and lists the biggest unsolved problems for researchers.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'first comprehensive survey' claim rests on an undocumented literature search; without a reproducible protocol, comprehensiveness and selection bias cannot be assessed.","rationale":"The reader's weakest assumption—that the literature search is comprehensive and representative—is exactly the load-bearing concern for this paper's central claim. Without a documented query, database list, and inclusion criteria, the assertion of being the first comprehensive survey is not independently checkable. The reader's CONDITIONAL verdict is appropriate: the survey is useful and the math appears sound, but the novelty and completeness claims require the authors to supply a reproducible search protocol. I find no additional internal inconsistency or mathematical error that would call for a stronger verdict, nor is there evidence to reject the paper. The concrete test would settle the concern by comparing the reference list against a systematic retrieve-then-filter search and by checking for prior surveys.","tokens_in":12276,"tokens_out":4607,"duration_ms":46604,"concrete_test":"Run a systematic search on DBLP, arXiv, and Scopus with the query ('spatiotemporal point process' AND ('neural' OR 'deep learning')) over 2017–2025, applying explicit inclusion criteria (papers proposing or applying neural STPPs). Compare the retrieved set with the survey's references and timeline; meanwhile, search for ('survey' OR 'review') AND 'spatiotemporal point process' AND ('neural' OR 'deep learning') to detect any prior comprehensive survey. If the survey omits more than about 10% of retrieved 2023–2025 primary papers, or if a prior survey with the same scope is found, the 'first comprehensive review' claim is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (Section 1: 'To our knowledge, no prior survey has comprehensively examined these aspects') is only as strong as the completeness and representativeness of the literature review. The search methodology is described in one sentence—'keyword-based queries, citation tracking, and seminal works in neural temporal point processes'—with no databases, query terms, date ranges, inclusion/exclusion criteria, or PRISMA-style flow. This makes the uniqueness assertion unverifiable and leaves open the possibility that relevant 2023–2025 work or a prior survey was missed. Many of the reviewed methods and applications are authored by the same research groups (e.g., Dong, Xie, Zhu, Okawa, and co-authors), which raises a legitimate selection-bias concern: the review may over-represent work the authors know well rather than the full field. This is a correctness risk for the central claim, not a reflection on the authors' integrity. Because the claim of being first and comprehensive is the main contribution, a reader cannot confirm it from the manuscript alone; a documented and reproducible search is the standard remedy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":12413,"tokens_out":3152,"duration_ms":34866,"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":[{"comment":"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.","section":"1 (Introduction, 'Scope and structure')"},{"comment":"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.","section":"1 (Introduction) and throughout"},{"comment":"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.","section":"3.2 (Single Event Prediction, Eq. (3))"}],"minor_comments":[{"comment":"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.","section":"5 (Evaluation Metrics)"},{"comment":"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.","section":"3.2 (Single Event Prediction, discussion of Li et al. [2024])"},{"comment":"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.","section":"7 (Open Challenges)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a useful survey, but its main claim of being the first comprehensive review depends on the completeness and representativeness of the literature search, which is currently undocumented. The heavy presence of works from a small set of research groups is a legitimate concern that should be addressed by the authors rather than left implicit. I would not recommend rejection, because the issues are fixable within the manuscript's scope: adding a documented search protocol, a coverage table, and a more cautious framing of the novelty claim would substantially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid survey of neural spatiotemporal point processes, and honestly, the field needed a map like this. The taxonomy of design choices—spatial encodings, kernel forms, history encoders, training objectives—is well organized and will help researchers orient themselves. The generalized kernel decomposition in Eq. (4) is a nice unifying notation that subsumes several existing approaches, even if it is not a new method. The mathematical background is standard and accurate, and the open challenges section names real problems: missing standardized benchmarks, limited reproducibility, and the gap between predictive accuracy and policy-relevant interpretability. The paper does what a good survey should do: it collects, organizes, and points at the hard parts.\n\nThat said, the central claim to be the first comprehensive survey is the weak spot. The literature search is described in one sentence—'keyword-based queries, citation tracking, and seminal works'—with no databases, query terms, date ranges, or inclusion criteria. That makes the comprehensiveness claim unverifiable, and the stress-test note is right about this. The concern about self-citation also has some force: a large share of the canonical examples come from a small set of groups (Mohler, Xie, Zhu, Dong, Okawa), so the review may over-represent work the authors know well. It is not a fatal flaw, and the paper does cite outside that circle (Lüdke, Yuan, Zhou, Erfanian, Jin, Zhang), but a documented search protocol plus a short discussion of selection bias would substantially strengthen the contribution.\n\nMinor issues: there is a duplicated sentence about 'Prediction accuracy (ACC)' in Section 5, and Eq. (3)'s explanatory text has a subscript slip (s vs s′). Copyediting stuff, not substance.\n\nWho is this for? New grad students and researchers entering neural STPPs. It gives a clear entry point and a reasonable bibliography. It is not a technical breakthrough, and it does not need to be. My verdict: the paper deserves a serious referee, but the authors should be asked to document the search methodology, tone down the uniqueness claim to match what is actually verifiable, and fix the small errors before publication.","headline":"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.","tokens_in":12975,"tokens_out":914,"would_cite":true,"duration_ms":11193,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["neural spatiotemporal point processes","conditional intensity function","deep learning survey","event prediction","Hawkes processes","kernel methods","diffusion models","uncertainty quantification"],"falsifier":"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.","tokens_in":12050,"feed_emoji":"📍","tokens_out":4684,"duration_ms":46928,"temperature":0.7,"pith_summary":"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.","feed_headline":"Survey maps neural spatiotemporal point processes","feed_subtitle":"A first comprehensive review unifies kernel, attention, and diffusion designs and pinpoints the field's missing benchmarks.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Establishes the prior art on neural temporal point process reviews, which the paper identifies as not covering spatial components.","marker":"[Shchur et al., 2021]"},{"why":"Represents the traditional statistical review of self-exciting STPPs that the paper contrasts with neural approaches.","marker":"[Reinhart, 2018]"},{"why":"Provides the continuous-time normalizing flow formulation that the paper presents as a foundational neural STPP method.","marker":"[Chen et al., 2021]"},{"why":"Contributes the Gaussian-mixture non-stationary kernel and imitation-learning objective that anchor the kernel and training discussions.","marker":"[Zhu et al., 2021b]"},{"why":"Introduces contextual deep mixture point processes that motivate the context-aware kernel line in the review.","marker":"[Okawa et al., 2019]"},{"why":"Supplies the diffusion-based joint spatiotemporal model used to illustrate intensity-free generative approaches.","marker":"[Yuan et al., 2023]"},{"why":"Offers the nonparametric kernel-plus-transformer model that the paper uses for kernel-based likelihood with amortized variational inference.","marker":"[Zhou et al., 2022]"},{"why":"Introduces automatic integration for spatiotemporal neural point processes, a key training innovation highlighted in the review.","marker":"[Zhou and Yu, 2024]"}],"fun_headline_variants":["Neural spatiotemporal point processes unified in one frame","Survey unifies neural STPP designs, flags missing benchmarks","Trends and challenges in neural spatiotemporal point processes","How neural point processes map space-time events","A roadmap for neural spatiotemporal point process research"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Neural spatiotemporal point processes unified in one frame","Survey unifies neural STPP designs, flags missing benchmarks","Trends and challenges in neural spatiotemporal point processes","How neural point processes map space-time events","A roadmap for neural spatiotemporal point process research"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000222,"raw_usage":{"total_tokens":1402,"prompt_tokens":839,"completion_tokens":563,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":489}},"tokens_in":455,"tokens_out":563,"duration_ms":5876,"temperature":1.0,"reasoning_tokens":489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:48:06.919161+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}