acceptodds
Under review as a conference paper at ICLR 2027

Learning a Scalable Neural Surrogate for Event-Level Statistically Grounded Hotspot Detection in Spatial Networks

Abstract

The growing availability of large-scale spatial event data has made hotspot detection increasingly challenging in spatial data analytics, with applications in public health, public safety, and urban analytics. In these domains, statistically grounded methods are widely used but computationally expensive when applied to millions of events on road networks. In this work, we propose a graph neural network framework that learns to approximate a computationally expensive statistical hotspot-detection method on road networks. Specifically, we use a network local K-function-based method to generate supervision, including event-level hotspot labels and associated hotspot radii, and train a graph neural network to jointly predict hotspot membership and radius directly from graph-structured spatial data. We construct heterogeneous graphs of events and road intersections, allowing the model to use both event features and road-network structure. By replacing repeated statistical evaluation at inference time, the learned surrogate accelerates event-level hotspot detection while maintaining good agreement with the statistical detector. We evaluate the surrogate on real-world road networks. On the Los Angeles test graph containing 8 million events, it achieves an F1 score of 0.897 and AP of 0.979. It processes the graph in 0.427 seconds, achieving an approximately 3,400-fold speedup over AH-IBT, the current state-of-the-art method. Radius predictions on the same graph have a mean absolute error of 22.1 meters. The model also transfers zero-shot to unseen cities.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Inductive Learning, Surrogate Modeling, Spatial Data Mining

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WLsBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a graph-based neural network for hotspot detection. Once trained, the model can be used as a surrogate for conventional hotspot detection methods such as the local K-function. The major advantages are its computational efficiency and zero-shot generalisation capability, e.g., a model trained in one city can be applied to other unseen cities as well.

Critical strengths
  • The authors incorporate road network information into spatial event datasets for hotspot detection and propose a new graph-based approach to model the relationships between events and road networks.
  • Manually designed node features are extracted to facilitate hotspot detection.
  • Once trained, the model is much more efficient than conventional hotspot detection methods, but with a trade-off in detection accuracy.
Critical weaknesses
  • While spatial hotspot detection is a well-established research problem, this paper uses an existing spatial hotspot detection method, i.e., the local K-function, to derive event-level hotspot labels and associated radii, and uses the derived information as ground truth to train a graph neural network. While this evaluation is appropriate for measuring surrogate approximation fidelity, it does not establish the actual hotspot detection accuracy or statistical reliability of the proposed method. Thus, the main advantage of this method is computational efficiency. However, do practical applications require real-time processing at the cost of reduced detection accuracy? Can the authors give a few examples where efficiency is a critical requirement in real-world scenarios?

  • I believe the original spatial event datasets do not include road network information. It seems that the authors enrich the datasets with road network information. However, it is unclear to me what the data source of the road network is. Is it from OpenStreetMap or other sources? What is the quality of the map data? It is possible that the road network has evolved over time. How do the authors ensure that the road network is correctly associated with historical events, considering missing roads or newly constructed roads? Is the method robust to errors in the map data? More analysis is required on these aspects.

  • The evaluation is not convincing. The authors only compare their method with a simple MLP baseline. This only shows that a graph-based neural network is able to learn additional information from graph structures, which is not surprising at all. The authors should compare their method with more advanced models, such as alternative graph neural networks or other deep learning based approaches.

  • There is also a lack of sufficient ablation studies for the proposed architecture. For example, how do accuracy and efficiency change with the number of GNN layers? There is insufficient verification of the contributions of each individual component, such as graph connectivity and handcrafted node features. Are there any other node features that could be incorporated to further enhance performance?

  • The generalization experiments demonstrate the model's ability to transfer across different cities. However, the event categories and their distributions in each dataset are not clearly specified. It is therefore unclear whether the cross-city experiments evaluate geographical generalization alone or also involve shifts in event types. The authors should clarify the event categories used in each dataset and conduct additional experiments to evaluate cross-event-type generalization, which is different from cross-city generalisation capabilities.

Questions for authors

Please refer to the questions listed in the weaknesses.

3 comments
AuthorsOpenReview ↗

WLsB: W1, W2

WLsB-W1

We thank the reviewer for this observation. We agree that our method is a surrogate for the local K-function-based AH-IBT detector, and that our evaluation measures fidelity to its outputs rather than independently establishing statistical significance guarantees.

Real-time applications of hotspot detection can trade accuracy in some specific circumstances in which detecting all hotspots offline is infeasible. For example, frequently updated applications, such as traffic congestion monitoring, service calls during natural disasters in populated regions, and thermal hotspots in energy systems, where repeatedly executing expensive statistical detection methods over millions of events delays updates, might prefer best effort detection rather than slow responses. Surges after storms or outbreaks also need quick updates. Each such query is a full statistical run over millions of events. On 8 million events, AH-IBT takes about 24.5 minutes end to end, whereas our model takes 8.4 seconds end to end.

When accuracy matters more, it can be recovered with extra computation: the optional calibration (Section 4.5) verifies uncertain predictions using AH-IBT, and its thresholds set how much is verified. On LA 8M, verifying part of the events brings recall and precision close to those of the statistical labels at a fraction of the cost of a full AH-IBT run.

Data

Table 17 (paper), largest graph per city, seed 42, thresholds (0.10, 0.95), two CPU cores.

GraphRecall (%)Precision (%)F1 (%)Verified (%)Time (s)
LA 8M95.6499.8697.731.84128.86
Chicago 7.68M94.0499.9696.928.0299.81
Detroit 5.17M94.2699.8297.038.0383.36
Seattle 214K98.4399.9499.233.249.23

For comparison on LA 8M: GraphSAGE-MT alone has recall 85.1%, precision 95.4%, F1 90.0%; with calibration, verification takes 128.9 s, about 11× less than a full AH-IBT run (1,464.6 s, excluding preprocessing).

WLsB-W2

The event data come from the cities' open-data portals cited in the paper, e.g., the Chicago Data Portal and the City of Los Angeles Open Data Catalog. Please refer to Appendix A.11 for the source, event type and time span of each dataset.

The road networks were obtained from OpenStreetMap (OSM) using OSMnx (Appendix A.11). We preprocess the networks by projecting coordinates, consolidating nearby intersections, removing parallel edges, and converting them to undirected graphs. Events are mapped to their nearest road segments using OSMnx's nearest_edges() function. Events on roads missing from the OSM snapshot (e.g., removed since 2010) are therefore not dropped but assigned to the nearest existing segment. Their positions along the edges are then calculated to support network-distance-based hotspot detection.

Please refer to Appendix A.9 for a robustness test with four types of map errors, each at three levels: missing roads, added roads, segment-length errors and event-position errors. Compared with the K-function detector rerun on the perturbed maps, F1 drops only slightly and only when roads are added, and GraphSAGE-MT's decisions change about as much as or less than the detector's own.

Data

Table 16 (paper), mean over LA 1M, Chicago 1M, Detroit 1M, Seattle 214K. (a) Against the K-function detector re-run on the perturbed map.

ErrorLevelF1 (%)MAPE
None—91.21.0%
Missing roads10%92.11.1%
Added roads10%89.51.1%
Length errorσ 20%91.21.1%
Position errorσ 50 m92.30.7%

(b) Against the labels of the clean map.

ErrorLevelF1 GraphSAGE (%)F1 K-function (%)Flips GraphSAGEFlips K-functionMAPE GraphSAGE
None—91.21000.0%0.0%1.0%
Missing roads10%90.495.25.9%6.0%9.7%
Added roads10%90.396.91.8%4.0%1.0%
Length errorσ 20%91.095.03.4%6.5%16.0%
Position errorσ 50 m86.189.514.1%13.2%52.7%

Flips = share of events whose hotspot decision differs from that on the clean map. GraphSAGE-MT flips fewer decisions than the detector under added roads and length errors, about as many under missing roads, and slightly more under position errors.

Highest level of each error shown. Full table (three levels each): Table 16 (paper, Appendix A.9).

Table 19 (paper).

CitySourceEvent typeTime spanEvents
LACity of Los Angeles open dataMixed city records1996–202517,641,139
ChicagoChicago Data Portal, CrimesCrime reports2001–20237,679,723
DetroitDetroit Open Data, 911 Calls for ServicePolice 911 calls2016–20225,171,281
SeattleSeattle SDOT CollisionsTraffic collisions2003–2020213,645
AuthorsOpenReview ↗

WLsB: W3, W4

WLsB-W3

We have added GCN, GAT and R-GCN as GNN baselines, trained on the same graph with the same features and protocol; please refer to Table 3 (Section 5.2) and Appendix A.7. R-GCN, which also distinguishes relation types, reaches nearly the same detection quality as GraphSAGE-MT, showing that our graph construction and features also work with other GNN architectures; GraphSAGE-MT uses less GPU memory, and GCN and GAT are weaker.

We have also added XGBoost as a stronger non-graph baseline, given the same information through precomputed neighborhood aggregates, to measure the contribution of message passing. It reaches nearly the same classification quality as GraphSAGE-MT but predicts radii less accurately. This demonstrates the effectiveness of our graph construction and features, which benefit even a non-graph model.

These baselines complement our main comparison with AH-IBT, the statistical detector our method aims to accelerate.

Data

Table 3 (paper): four-city mean, same inputs.

ModelF1Spec.AUROCAPMAPE
GraphSAGE-MT0.8950.8580.9530.9701.8%
R-GCN0.8940.8510.9520.9691.9%
GCN0.6620.7330.9310.9531.9%
GAT0.8590.7700.9240.9502.0%
XGBoost0.8870.8880.9530.9713.2%

Table 13 (paper): LA 8M, one 24 GB GPU, seed 42.

ModelForward (s)Peak memory (GiB)
GraphSAGE-MT0.72910.23
R-GCN1.08414.73
GCNOOMOOM
GATOOMOOM

Four-city mean. XGBoost-prop32 is the XGBoost of Table 3; XGBoost-8 drops the multi-hop aggregates.

ModelF1Prec.RecallSpec.AUROCAPMAPE
GraphSAGE, 4 layers0.8950.9040.8910.8590.9540.9701.9%
GraphSAGE, 1 layer0.870———0.928—1.8%
XGBoost-80.8820.8600.9150.7400.9490.9692.8%
XGBoost-prop320.8870.9200.8620.8880.9530.9713.2%

Without propagation, XGBoost specificity drops from 0.888 to 0.740.

WLsB-W4

Regarding GNN depth, accuracy improves up to three to four layers and then levels off, as each layer extends the receptive field by one hop along the graph, and forward time grows roughly linearly with depth.

Event features describe the road segment each event is on, which mainly determines the radius. Intersection features describe how many roads and events are around each intersection. The graph connections pass this information between nearby segments, so the model can tell whether events are concentrated over a wider area rather than on a single segment, which helps detect hotspot membership.

We have added a non-graph baseline, XGBoost, given the same input information through neighborhood aggregates precomputed on our graph (Table 3 and Appendix A.7); it reaches nearly the same classification quality as GraphSAGE-MT but predicts radii less accurately. We have also evaluated GraphSAGE without intersection–intersection edges: F1 is unchanged, but specificity drops, i.e., more false positives, showing the importance of connectivity between intersections.

We also zeroed each input feature in turn and retrained; the event-density rank of each event and the number of events around each intersection matter most, while each other feature has little effect. We have tried different features, including total distance, number of events, left and right distances for E-nodes, as well as average O–O and O–E distances and the numbers of E and O nodes for O-nodes. We also experimented with density relative to each graph’s mean density and raw event counts; neither improved on the current features overall.

Data

Depth, four-city mean; forward on LA 8M, seed 42.

LayersParamsF1AUROCMAPEF1 LAChicagoDetroitSeattleForward A10 (s)Forward 3090 (s)
110,4340.8700.9281.8%0.8410.8680.8810.8890.1750.126
359,9700.8950.9531.8%0.8760.8950.8880.9200.5430.458
484,7380.8940.9541.9%0.8770.8940.8850.9210.7290.573
6134,2740.8910.9521.9%0.8770.8930.8720.9201.0920.707

Layers 1, 3, 4, 6 shown.

Finer ablation, four layers, four-city mean (zeroed in training and testing).

VariantF1Prec.RecallSpec.AUROCMAPE
Full model0.8940.9040.8910.8590.9541.9%
No O–O edges0.8950.8890.9070.8250.9511.9%
Zero E1 (event-density rank)0.8800.8300.9480.6580.9471.7%
Zero O3 (#E neighbours rank)0.8410.9390.7720.9330.9451.8%
Zero all O features0.8440.9240.7860.9120.9411.9%

Each other single feature changes F1 by ≤ 0.005.

WLsB-W5

The datasets contain crime reports (Chicago), police 911 calls (Detroit), traffic collisions (Seattle) and mixed city records (LA, mainly inspections and permits). Please refer to Appendix A.11 for the categories and event distribution per city.

The cities record different types of events, so our cross-city tests involve a shift in event types as well as in geography. The model still transfers well to the unseen cities, which suggests that it generalizes across event types, not only across locations. Please refer to Table 4.

For cross-event-type generalization, please refer to Appendix A.8: the LA-trained model performs well on each of eight Chicago event types, and changing either the training data or the test type affects performance only slightly, showing that the model generalizes across event types.

Data

Table 19 (paper).

CitySourceEvent typeTime spanEvents
LACity of Los Angeles open dataMixed city records1996–202517,641,139
ChicagoChicago Data Portal, CrimesCrime reports2001–20237,679,723
DetroitDetroit Open Data, 911 Calls for ServicePolice 911 calls2016–20225,171,281
SeattleSeattle SDOT CollisionsTraffic collisions2003–2020213,645

Table 20 (paper): LA record types.

TypeTime spanPercentage
Inspections2013–202554%
Permits1996–202528%
Arrests2010–20259%
Crime reports2020–20245%
Traffic collisions2010–20253%
Service requests, COVID-192016–2023< 1%

Table 4 (paper): F1 on the full test graph, LA-trained model.

CityF1
LA (8M)0.899
Chicago0.891
Detroit0.886
Seattle0.939

Table 14 (paper): LA-trained model on single-type Chicago datasets (250K events each).

TypeHotspot (%)F1Prec.RecallSpec.AUROCMAPE
All types (control)55.90.9120.9310.8940.9170.9731.2%
Theft56.50.9340.9460.9220.9310.9821.0%
Criminal damage46.90.8690.8610.8780.8740.9551.5%

Hotspot (%) = percentage of hotspot events in each dataset.

Highest (theft) and lowest (criminal damage) of eight types shown; F1 0.869–0.934. Full table: Table 14 (paper, Appendix A.8).

Table 15 (paper): F1, rows = training data, columns = Chicago test type.

Trained onAllTheftBatteryCrim. dam.NarcoticsAssaultBurglaryMV theftRobbery
Chicago theft0.9200.9430.9260.8730.9410.9220.8800.8810.936
Chicago battery0.9230.9240.9460.8750.9670.9380.8840.8820.945
Chicago all types0.9300.9380.9390.8850.9610.9340.8920.8920.946
LA0.9120.9340.9120.8690.9260.9090.8720.8750.924

Reading down each column, F1 changes by at most 0.041 when the training data change; reading across each row, it changes by 0.065–0.092 between test types. Both effects are small, showing that the model generalizes across event types.

Reviewer_aEaQOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To address the prohibitive computational cost of statistically grounded hotspot detection on road networks with millions of events, this paper proposes a GNN surrogate that learns to approximate the statistical detector: supervision, consisting of event-level hotspot labels and associated radii, is generated offline by the statistical detector, and a multi-task GraphSAGE is trained on a heterogeneous intersection–event graph to jointly predict hotspot membership and radius for each event, complemented by an optional calibration step with selective statistical verification. On the 8-million-event Los Angeles graph, the model completes inference in 0.427 seconds, achieving agreement with the statistical labels of F1 = 0.897 and AP = 0.979 with a radius MAE of 22.1 meters, and transfers zero-shot to three unseen cities.

Critical strengths

(1) Targeting the practical challenges of high computational cost and limited scalability of statistical hotspot detection on large-scale spatial event data, this paper replaces the original statistical detection pipeline with a learned surrogate that supports zero-shot cross-city transfer, offering clear practical value for applications in public safety, public health, and urban planning.

Critical weaknesses

(1) Severely insufficient baselines: MLP (no graph) is the only learning-based baseline. Comparisons with other traditional machine learning and deep learning methods are missing, which is insufficient to substantiate the effectiveness of the proposed method.

(2) Limited experimental analysis and result presentation: results are reported only as aggregate numerical metrics, without error analysis (the spatial distribution of false positives/negatives, the statistical properties of radius errors) or any visualization of the detection results, making it difficult to intuitively assess the spatial plausibility of the outputs.

Questions for authors

(1) The exposition of the core concept of "hotspot detection" is not sufficiently clear: please define it more precisely—what events does a "hotspot" specifically refer to in this paper?

(2) The dataset descriptions lack clarity: what are the time spans of the four city datasets, the composition of event types (the proportion of crime records vs. service calls), how events are mapped to road segments, and the exact rules for splitting the 16-million-event LA pool into training/validation/test (is there any spatio-temporal overlap between training and test)?

(3) The experiments include only one learning-based baseline, MLP (no graph), which is insufficient to support conclusions about the method's effectiveness. Why are other traditional machine learning and deep learning baselines missing—for example, XGBoost/GBDT built on multi-scale neighborhood count features (the K-function is essentially such a feature), or alternative GNN architectures such as GCN/GAT?

(4) Compared against the only learning baseline, MLP, the improvement in classification metrics is modest (F1: 0.806 → 0.897), and the MLP itself is a weak baseline. How can you demonstrate that the gains from graph-based modeling are significant and necessary? Could this be verified with a stronger non-graph baseline?

(5) The paper discloses that the reported GNN forward time excludes graph construction and data transfer, but does the timing of AH-IBT and the exact K-function include preprocessing steps such as index construction? Furthermore, supervision for the 500K training events must be generated by the statistical method before training—what is the cost of this one-time offline labeling, and of training itself?

(6) Appendix A.3 ablates feature groups, but on what basis were the remaining design choices determined—e.g., the edge-type definitions and the feature transformations (log(1+x), within-graph percentile ranks)? What is the rationale for α = 0.94?

(7) As application-oriented work on road networks, would it be possible to visualize the detection results on the road network (e.g., annotating hotspot membership and predicted radii, overlaid against the statistical detector's output), so as to intuitively convey the spatial patterns of the outputs and the spatial distribution of errors?

5 comments
AuthorsOpenReview ↗

aEaQ: W1, W2, Q1

aEaQ-W1

We have added two groups of baselines: other GNN architectures (GCN, GAT and R-GCN), and a non-graph XGBoost model that takes the same features averaged over up to three intersection hops; please refer to Table 3 (Section 5.2) and Appendix A.7. Relation-aware models (GraphSAGE-MT, R-GCN and XGBoost) reach nearly the same AUROC and AP, so the gain comes mainly from our graph and feature design.

Data

Table 3 (paper): four-city mean, same inputs.

ModelF1Spec.AUROCAPMAPE
GraphSAGE-MT0.8950.8580.9530.9701.8%
R-GCN0.8940.8510.9520.9691.9%
GCN0.6620.7330.9310.9531.9%
GAT0.8590.7700.9240.9502.0%
XGBoost0.8870.8880.9530.9713.2%

Table 13 (paper): LA 8M, one 24 GB GPU, seed 42.

ModelForward (s)Peak memory (GiB)
GraphSAGE-MT0.72910.23
R-GCN1.08414.73
GCNOOMOOM
GATOOMOOM

aEaQ-W2

We have added error analysis and visualization; please refer to Figure 3 (Section 5.2), a Seattle road-network map comparing the hotspots and radii of our model and the K-function detector, with false positives and false negatives marked, and to Appendix A.4 for recall and radius error by ground-truth radius. Most errors lie on hotspot boundaries, and hotspots missed entirely are rare.

Data

Seattle, seed 42 (Figure 3).

EventsErrorsError rate
Boundary events52%97%15.9%
Other events48%3%0.6%

A boundary event is an event on a road segment where the segment and its adjacent segments contain both hotspot and non-hotspot events.

Hotspots missed entirely hold 1% of hotspot events.

Table 8 (paper): groups = quintiles of the ground-truth radius over LA training hotspots.

Radius (m)Recall LAChicagoDetroitSeattleMAPE LAChicagoDetroitSeattle
All0.9100.8840.8520.9231.5%1.8%1.3%2.7%
≤ 1140.9170.9040.9060.9661.9%1.8%1.3%2.4%
114–1810.9360.9120.9220.9462.0%2.9%1.8%4.1%
181–2230.9180.8720.7700.8871.2%1.1%1.3%2.1%
223–3470.8920.8780.6990.8951.4%2.1%1.1%2.5%
> 3470.8830.8060.7900.6710.8%0.5%0.4%1.2%

Hotspots are split into five equal-size groups by ground-truth radius over LA training hotspots. Radius MAPE stays at or below 4.1% in every group. Recall is lower in the larger-radius groups, most clearly in Seattle above 347 m (0.671). Large radii mostly come from long segments, whose lower event density ranks low within cities where such hotspots are rare, so they are more often missed.

aEaQ-Q1

Please refer to the detailed problem formulation in Section 3: a hotspot is a place on the road network where events, such as crimes, service calls or traffic collisions, are significantly denser than chance. A place is a hotspot if the number of events within network distance t around it exceeds the CSR null at level α (K(t|ϑ) > Kα(t|ϑ)); its radius is where this excess peaks. We have stated this more clearly and added Figure 1 with an explanatory paragraph in Section 1.

AuthorsOpenReview ↗

aEaQ: Q2, Q3, Q4

aEaQ-Q2

Please refer to Appendix A.11 (Tables 19–20) for the source, event types and time span of each dataset. LA combines several city record types (mainly inspections and permits, plus arrests, crime reports and traffic collisions) from 1996 to 2025, 89% from 2010 or later. GraphSAGE performance on different event types is reported in Appendix A.8.

As for event mapping, each event is snapped by its coordinates to the nearest road segment. LA is split at random with a fixed seed into a 1M training pool (500K training and 500K validation) and a 16M test pool, so training, validation and test share no events. These details are in Appendix A.11.

Spatial overlap does not leak labels: the model takes no coordinates or node IDs, and each test graph’s features and hotspot labels are computed only from its own events; on three unseen cities with no shared roads, F1 is still 0.886–0.939. Time is neither a model input nor part of the hotspot definition (the network K-function is purely spatial), so temporal overlap does not affect the task.

Data

Table 19 (paper).

CitySourceEvent typeTime spanEvents
LACity of Los Angeles open dataMixed city records1996–202517,641,139
ChicagoChicago Data Portal, CrimesCrime reports2001–20237,679,723
DetroitDetroit Open Data, 911 Calls for ServicePolice 911 calls2016–20225,171,281
SeattleSeattle SDOT CollisionsTraffic collisions2003–2020213,645

Table 20 (paper): LA record types.

TypeTime spanShare
Inspections2013–202554%
Permits1996–202528%
Arrests2010–20259%
Crime reports2020–20245%
Traffic collisions2010–20253%
Service requests, COVID-192016–2023< 1%

Table 4 (paper): F1 on the full test graph, LA-trained model.

CityF1
LA (8M)0.899
Chicago0.891
Detroit0.886
Seattle0.939

Table 14 (paper): LA-trained model on single-type Chicago datasets (250K events each).

TypeHotspot (%)F1Prec.RecallSpec.AUROCMAPE
All types (control)55.90.9120.9310.8940.9170.9731.2%
Theft56.50.9340.9460.9220.9310.9821.0%
Battery63.50.9120.9630.8660.9420.9741.2%
Criminal damage46.90.8690.8610.8780.8740.9551.5%
Narcotics74.00.9260.9910.8690.9780.9861.0%
Assault62.00.9090.9510.8710.9270.9701.2%
Burglary46.90.8720.8740.8710.8880.9561.5%
Motor vehicle theft49.60.8750.8790.8710.8820.9551.5%
Robbery65.30.9240.9680.8840.9450.9781.1%

aEaQ-Q3

We have added GCN, GAT and R-GCN as GNN baselines, and XGBoost as a non-graph baseline; please refer to Table 3 (Section 5.2) and Appendix A.7. The GNN baselines use the same graph and node features, while XGBoost uses the same features averaged over up to three intersection hops. Relation-aware models (GraphSAGE-MT, R-GCN and XGBoost) reach nearly the same AUROC and AP, showing that our graph construction and features also work with other architectures.

These baselines complement our main comparison with AH-IBT, the statistical detector our method aims to accelerate.

Data

Table 3 (paper): see aEaQ-W1.

aEaQ-Q4

The main benefit comes from the neighborhood information on our graph: propagating features over the graph improves precision and specificity, even for a non-graph model. GNNs learn this end to end, while non-graph models must precompute it.

To support this empirically, we compare GraphSAGE with two baselines. One is XGBoost (non-graph baseline), given the same features averaged over up to three intersection hops on our graph (Table 3, Appendix A.7). It reaches nearly the same AUROC and AP as GraphSAGE-MT, and without this propagation its specificity drops from 0.888 to 0.740, so our graph construction also benefits non-graph models. The other is a one-layer GraphSAGE, which removes only wider-neighborhood information and has lower F1 and AUROC than the four-layer model (0.870 / 0.928 vs. 0.895 / 0.954).

Data

Four-city mean. XGBoost-prop32 is the XGBoost of Table 3; XGBoost-8 drops the multi-hop aggregates.

ModelF1Prec.RecallSpec.AUROCAPMAPE
GraphSAGE, 4 layers0.8950.9040.8910.8590.9540.9701.9%
XGBoost-80.8820.8600.9150.7400.9490.9692.8%
XGBoost-prop320.8870.9200.8620.8880.9530.9713.2%

GraphSAGE depth, four-city mean.

LayersF1AUROCMAPE
10.8700.9281.8%
40.8950.9541.9%
AuthorsOpenReview ↗

aEaQ: Q5, Q6, Q7

aEaQ-Q5

The reported AH-IBT and K-function runtimes exclude reading the raw input file and constructing their input graphs. We have measured this preprocessing and report end-to-end times in Section 5.2.

Training has three one-time costs on LA500K: generating the labels for the training and validation splits takes about 290 s each on a CPU, graph construction takes under 1 s per split, and training takes 45.5 s on an RTX 3090, about 11 minutes in total.

After that, inference requires no labeling. On LA 8M, end-to-end inference from the raw file on one RTX 3090 takes 8.4 s (7.6 s of graph construction), whereas AH-IBT and the exact K-function take about 1,470 s and 11,140 s including the construction of their input graphs, so GraphSAGE-MT is about 180× and 1,300× faster. The trained model is reused for every test graph and every unseen city. Graph construction takes 0.1–1.1 s for test graphs of up to 1M events and under 8 s for the full 5–8M-event graphs.

Data

LA 8M end to end, from the raw file (paper, Section 5.2; RTX 3090).

StageGraphSAGE-MT (s)
Graph construction7.56
Inference0.79
End to end8.4
MethodExcl. preprocessing (s)Preprocessing (s)End to end (s)GraphSAGE-MT faster
AH-IBT1,464.588.25≈ 1,470≈ 180×
K-function11,099.1743.86≈ 11,140≈ 1,300×

Training cost (LA500K).

StageTime (s)
Labels, train291.1
Labels, validation≈ 295
Graph construction, train + validation0.74 + 0.74
Training, 500 epochs, RTX 3090 (mean of seeds 42/43/44: 45.3, 45.6, 45.7)45.5
Total≈ 633 (≈ 11 min)

Graph construction per test graph (RTX 3090 node, 1 CPU).

Test graphEventsTime (s)
LA20K0.25
LA100K0.39
Seattle (full)214K0.27
LA1M1.14
Detroit (full)5.17M4.19
Chicago (full)7.68M6.43
LA8M6.78

aEaQ-Q6

O–O edges keep the road topology, so information travels along roads as hotspots do; removing them lowers specificity from 0.859 to 0.825.

Each event is linked only to its segment’s two endpoints, so events on the same or nearby segments exchange information through intersections, and the number of edges grows linearly with the number of events instead of quadratically, as direct event–event edges would.

log(1+x) compresses heavy-tailed distances and lengths while keeping zero defined. Within-graph percentile ranks put densities on the same scale in every city, so a model trained on LA can be applied to other cities without rescaling.

α = 0.94 follows the setting of Liu et al. (2023), whose detectors we use for labeling and timing.

Appendix A.1 notes the alternative density features we tried, density relative to each graph's mean density and raw event counts, which did not improve on the current features overall.

Data

Four layers, four-city mean.

VariantF1Prec.RecallSpec.AUROCMAPE
Full model0.8940.9040.8910.8590.9541.9%
No O–O edges0.8950.8890.9070.8250.9511.9%

aEaQ-Q7

Yes. We have added a road-network map figure (Figure 3, Section 5.2) comparing our hotspots and radii with the K-function detector on Seattle, with false positives and false negatives marked.

Data

Figure 3: Seattle (213,645 events), seed 42; map window 2.4 km × 2.4 km with 15,359 events.

EventsErrorsError rate
Boundary events52%97%15.9%
Other events48%3%0.6%

Hotspots missed entirely hold 1% of hotspot events.

Reviewer_Y46nOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses statistical hotspot detection in large-scale road networks and proposes GraphSAGE-MT as a learning-based surrogate for a network local K-function-based statistical detector. The method constructs a sparse heterogeneous graph containing event nodes and road intersection nodes, and uses a shared GraphSAGE backbone to jointly predict event-level hotspot membership and the corresponding radius. The authors also introduce an optional statistical calibration procedure to selectively verify hotspot predictions.

The model is trained and validated on Los Angeles data and evaluated on the Los Angeles test set and three unseen cities. The results show good agreement with the statistical detector's outputs, short GPU forward-pass times, and encouraging zero-shot transfer performance across cities.

Critical strengths
  1. The research problem has clear practical relevance, and the method design considers the scalability requirements of the task. The paper addresses the computational cost of statistical hotspot detection on large-scale road networks. It uses hotspot labels and radii generated by a statistical detector to train a learning-based surrogate that predicts both hotspot membership and spatial extent. The proposed sparse heterogeneous graph connects event nodes through road intersections, avoiding potentially expensive dense event-to-event connections. This design is relevant to processing large-scale spatial event data and has potential practical value.

  2. The experiments cover large-scale datasets and include cross-city evaluation. The paper evaluates prediction accuracy and computational efficiency on a Los Angeles test graph containing 8 million events, and conducts zero-shot evaluations on three unseen cities: Chicago, Detroit, and Seattle. The reported results show that GraphSAGE-MT outperforms the MLP baseline in hotspot classification and radius prediction, while achieving short GPU forward-pass times on large graphs. The paper also includes feature ablation experiments and evaluations of the optional statistical calibration procedure, providing additional evidence of the model's performance.

Critical weaknesses
  1. The contribution of graph-based modeling to the observed performance gains could be further clarified. Section 5.1 compares the predictive performance of GraphSAGE-MT with an MLP baseline, while Table 7 (Appendix A.3) further examines the effects of event-node and intersection-node features. These experiments provide useful evidence for the effectiveness of the model design. However, GraphSAGE-MT and the MLP differ in both input information and parameter count (Table 5, Appendix A.1), so the observed performance gap may reflect several factors at once. Since heterogeneous graph modeling is a central design choice of this work, further discussion of how the existing results support the contribution of graph structure and message passing would be helpful.

  2. The joint evaluation of hotspot membership and predicted radius could be further strengthened. The paper evaluates the model's two outputs separately, using classification metrics for hotspot membership and prediction errors for the radius. The radius errors are computed on events labeled as hotspots by the statistical detector (Section 5.1). In addition, the optional calibration procedure in Section 4.5 selectively applies statistical verification to hotspot predictions while retaining the model-predicted radii. These evaluations provide useful evidence of the predictive performance on each task. However, the current results do not directly show whether events predicted as hotspots also satisfy the original detector's significance criterion at their predicted radii.

Questions for authors

Q1. Relationship between hotspot membership and predicted radius For events predicted as hotspots by the model, could the authors clarify whether the predicted radius generally satisfies the significance condition of the original statistical detector? For example, could the authors report, on a subset of test events, the proportion of predicted hotspots satisfying K(r^∣ϑ)>Kα(r^∣ϑ)K(\hat{r} \mid \vartheta) > K_\alpha(\hat{r} \mid \vartheta)? This would complement the classification metrics and radius prediction errors that are currently reported separately.

Q2. Network distance information and different radius scales The message-passing formulation in Section 4.4 does not explicitly incorporate road-edge lengths. Could the authors further explain how the model uses the distance-related features to approximate statistical detection across different network-distance scales? In addition, could the authors group events by their ground-truth hotspot radius and report the hotspot recall and radius prediction error for each group? This would help clarify how the model performs across different spatial scales.

Q3. Comparability of the baseline In Section 5.1, the MLP and GraphSAGE-MT differ in both input information and parameter count (Table 5, Appendix A.1). Although Table 7 provides feature ablations, could the authors further clarify how the existing experiments distinguish the benefits of additional neighborhood information from those of graph message passing? If feasible, a non-graph baseline with more comparable input information and model capacity would help further clarify the contribution of graph-based modeling.

Q4. Role of joint training Have the authors compared the current setup with separately trained models for hotspot classification and radius regression, or examined how different values of the loss weight λ\lambda affect performance on the two tasks? If such experiments are not available, could the authors briefly explain the rationale for the current joint-training design and the choice of λ=1.0\lambda = 1.0?

Q5. Classification evaluation under different class ratios Table 2 shows that hotspot prevalence varies from approximately 60% to 74% across the four cities. Although the paper already reports F1 and AP, differences in class prevalence may affect their interpretation, particularly for AP. Could the authors consider additionally reporting AUROC as a complementary threshold-independent metric to facilitate the comparison of classification performance across cities?

3 comments
AuthorsOpenReview ↗

Y46n: W1, W2&Q1, Q2

Y46n-W1

In Table 7, after removing intersection features, the results show how graph-based neighbor information improves performance. We have added a non-graph baseline, XGBoost, given the same input information through neighborhood aggregates precomputed on our graph, and applied our graph construction to GCN, GAT and R-GCN; please refer to Table 3 (Section 5.2) and Appendix A.7. Relation-aware models, including XGBoost, reach nearly the same AUROC and AP, so the gain comes mainly from our graph and feature design and carries over to non-graph models instead of certain model architecture. GraphSAGE-MT predicts radii more accurately than XGBoost.

Data

Table 3 (paper): four-city mean, same inputs.

ModelF1Spec.AUROCAPMAPE
GraphSAGE-MT0.8950.8580.9530.9701.8%
R-GCN0.8940.8510.9520.9691.9%
GCN0.6620.7330.9310.9531.9%
GAT0.8590.7700.9240.9502.0%
XGBoost0.8870.8880.9530.9713.2%

Table 7 (paper): LA, GraphSAGE-MT retrained with intersection or event features set to zero.

ModelF1Prec.RecallSpec.APAUROCMAPE
Full0.8770.8530.9110.8450.9590.9581.5%
without intersection0.8590.8540.8670.8550.9440.9411.5%
without event0.8610.8240.9200.7950.9530.95425.2%

Both feature types carry event density, so message passing partly makes up for the zeroed type. Without intersection features, recall drops. Without event features, events on the same segment get identical inputs, so specificity drops and radius MAPE rises to 25.2%.

Y46n-W2 and Y46n-Q1

Yes for true positives: across all cities, 99.3–99.8% of true positives satisfy the significance condition at r^\hat{r}, while 5.4–47.8% of false positives do.

Please refer to Appendix A.5 (Tables 9–10) for the pass rate per city and per test graph, split by true and false positives.

Data

Table 9 (paper): K-function test at the predicted radius r̂.

CityTestedTP percentage (%)TP pass rate (%)FP percentage (%)FP pass rate (%)
LA334,14186.799.313.35.4
Chicago333,58492.399.67.710.8
Detroit337,28893.499.86.647.8
Seattle182,37693.299.76.822.5

Y46n-Q2

Distances enter through node features, not edge weights. Event nodes carry their segment length and distances to both endpoints; intersection nodes carry average distances to neighboring intersections and events. Each layer adds one more hop, so deeper layers cover larger network distances. With one layer, four-city F1 is 0.870; two to four layers reach 0.893–0.895.

For the ground-truth grouping, please refer to Appendix A.4 (Table 8), which reports recall and radius MAPE by ground-truth radius quintile, per city. Radius error stays low in every group, and recall is lower in the larger-radius groups.

Data

Depth, four-city mean.

LayersParametersF1AUROCMAPE
110,4340.8700.9281.8%
235,2020.8930.9491.7%
359,9700.8950.9531.8%
484,7380.8940.9541.9%

Table 8 (paper): groups = quintiles of the ground-truth radius over LA training hotspots.

Radius (m)Recall LAChicagoDetroitSeattleMAPE LAChicagoDetroitSeattle
All0.9100.8840.8520.9231.5%1.8%1.3%2.7%
≤ 1140.9170.9040.9060.9661.9%1.8%1.3%2.4%
114–1810.9360.9120.9220.9462.0%2.9%1.8%4.1%
181–2230.9180.8720.7700.8871.2%1.1%1.3%2.1%
223–3470.8920.8780.6990.8951.4%2.1%1.1%2.5%
> 3470.8830.8060.7900.6710.8%0.5%0.4%1.2%

Hotspots are split into five equal-size groups by ground-truth radius over LA training hotspots. Radius MAPE stays at or below 4.1% in every group. Recall is lower in the larger-radius groups, most clearly in Seattle above 347 m (0.671). Large radii mostly come from long segments, whose lower event density ranks low within cities where such hotspots are rare, so they are more often missed.

AuthorsOpenReview ↗

Y46n: Q3, Q4, Q5

Y46n-Q3

The main benefit of message passing is learning neighborhood information end to end. Our heterogeneous graph builds this information into intersection features, and message passing lets GNN gather it directly from the graph, while a non-graph model must precompute it.

To support this, we compare GraphSAGE with two baselines. One is XGBoost given the same graph-derived features, including neighborhood aggregates precomputed on our graph. Removing these features would give XGBoost less information and make the comparison unfair. Keeping them removes only message passing, and also shows that our graph construction benefits non-graph models. The other is a one-layer GraphSAGE, which removes only wider-neighborhood information and shows its importance. Please refer to Table 3 (Section 5.2) and Appendix A.7 for the XGBoost comparison.

Data

Table 3 (paper): four-city mean, same inputs.

ModelF1Spec.AUROCAPMAPE
GraphSAGE-MT0.8950.8580.9530.9701.8%
R-GCN0.8940.8510.9520.9691.9%
GCN0.6620.7330.9310.9531.9%
GAT0.8590.7700.9240.9502.0%
XGBoost0.8870.8880.9530.9713.2%

Four-city mean. XGBoost-prop32 (the XGBoost in Table 3) uses the same input features as GraphSAGE and a similar number of parameters; XGBoost-8 drops the multi-hop aggregates.

ModelF1Prec.RecallSpec.AUROCAPMAPE
GraphSAGE, 4 layers0.8940.9040.8910.8590.9540.9701.9%
GraphSAGE, 1 layer0.8700.8630.8840.7620.9280.9531.8%
XGBoost-80.8820.8600.9150.7400.9490.9692.8%
XGBoost-prop320.8870.9200.8620.8880.9530.9713.2%

Without message passing, XGBoost-prop32 matches GraphSAGE in AUROC and AP but has higher radius MAPE (3.2% vs. 1.9%). Without wider-neighborhood information (XGBoost-8, one-layer GraphSAGE), specificity drops to 0.740–0.762.

Y46n-Q4

We compared joint training with separately trained models and swept λ∈{0.1,0.3,1,3,10}\lambda \in \{0.1, 0.3, 1, 3, 10\} (three seeds each). Please refer to Appendix A.6 (Table 11).

A separately trained classifier matches the joint model. Joint training also lets both tasks share one network and one forward pass, thus saving time.

Data

Table 11 (paper).

λLA 8M F1LA 8M MAPEFour-city F1Spec.AUROCMAPE
0.10.9000.7%0.8970.8500.9541.8%
0.30.9000.7%0.8970.8500.9541.8%
10.8990.7%0.8950.8580.9531.8%
30.8970.7%0.8950.8580.9541.8%
100.8990.8%0.8960.8550.9541.9%

The results are not sensitive to λ.

Joint vs. separately trained classification, four-city mean.

ModelF1Spec.APAUROC
Joint0.8950.8580.9700.953
Classifier only0.8960.8540.9690.953

Joint vs. separately trained radius prediction, four-city mean.

ModelMAPE
Joint1.8%
Radius only1.8%

Separately trained models give the same classification (F1 0.896 vs. 0.895) and radius error (MAPE 1.8%) as the joint model, and it needs one model and one forward pass instead of two.

Y46n-Q5

We have added AUROC for all models and cities; please refer to Tables 3, 4 and 12. It measures how well the model distinguishes hotspots from non-hotspots in predictions, independent of hotspot ratio, and provides a common scale for comparing cities.

Data

GraphSAGE-MT AUROC (paper). Table 4 = full test graph; Table 12 = mean over the city's test graphs.

CityTable 4 (full graph)Table 12 (all test graphs)
LA0.9680.958
Chicago0.9750.966
Detroit0.9360.933
Seattle0.9750.957

Thank you for the detailed rebuttal and substantial additional experiments. The new baselines, significance checks, and ablation studies address most of my concerns. Although some questions about the distinct contribution of graph message passing remain, I am raising my rating from 2 to 3.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summarization for AC

We thank the AC and reviewers for their time and constructive feedback. We have addressed all the points raised by the reviewers and uploaded a revised version of our manuscript incorporating the new experiments, results, and corresponding discussions. We hope these revisions adequately address the reviewers’ concerns and further strengthen our work.

For convenience, we summarize the main revisions and findings below:

  1. Additional Baselines and ablation studies. We added four new competitive baselines (GCN, GAT, R-GCN, and non-graph XGBoost) to demonstrate the generalizability of our graph representation and feature design across different learning architectures. R-GCN and XGBoost achieve comparable AUROC and F1 scores to GraphSAGE-MT. This shows that our heterogeneous graph construction and feature design transfer to other models, including a non-graph one. However, GraphSAGE-MT remains our preferred architecture as it achieves the best F1 and radius prediction accuracy while directly processing the heterogeneous graph without requiring precomputed neighborhood aggregates. We also conducted additional ablation studies on GNN depth, graph connectivity, individual node features, joint training, and loss-weight sensitivity. (see Y46n-W1, Y46n-Q3–Q4, aEaQ-W1, WLsB-W3–W4).

  2. Finer-grained Evaluation. We also evaluated GraphSAGE-MT on unseen cities across different hotspot radius ranges. Radius prediction error remains low across all radii ranges, but recall decreases for larger radii. In Seattle, we achieve 67.1% recall for hotspots with radii greater than 347 m. However, this gap is largely recovered through calibration (Section 4.5). We achieve 88.6% and 98.5% recall while remaining approximately 9× and 4.5× faster than AH-IBT, respectively. Furthermore, in Seattle, 97% of errors lie on hotspot boundaries, where hotspot and non-hotspot events share the same or adjacent road segments, and hotspots missed entirely account for only 1% of hotspot events (see Y46n-Q2, aEaQ-W2).

  3. End-to-end efficiency and calibration. We extended our efficiency evaluation to account for graph construction, data transfer, offline labeling, and model training. The total one-time training cost is approximately 11 minutes. On LA 8M, GraphSAGE-MT remains approximately 180× faster than AH-IBT when accounting for preprocessing. Furthermore, our optional calibration procedure provides a flexible accuracy–efficiency trade-off, achieving 95.6% recall and 99.9% precision while remaining approximately 9-11× faster than AH-IBT (see aEaQ-Q5, WLsB-W1).

  4. Data, robustness, and generalization. We added a more detailed description of each dataset, including its source, event types, time span, and splitting procedure. We also conducted a robustness study covering four types of map errors at three severity levels, demonstrating that our model maintains high fidelity to the statistical detector under map perturbations. Furthermore, we evaluated cross-event-type generalization, where our LA-trained model achieves F1 scores ranging from 86.9% to 93.4% across eight Chicago event types without retraining (see aEaQ-Q2, WLsB-W2, WLsB-W5).

We hope these revisions address the reviewers’ concerns and further strengthen our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.