acceptodds
Under review as a conference paper at ICLR 2027

CSTP: Data-Efficient Road-Closure Prediction with Structured Inventory Dynamics

Abstract

Road-closure planning requires rapid predictions, even though comparable disruptions are relatively rare in real-world data. We present the Closure-aware Structured Traffic Predictor (CSTP), which predicts vehicle inventory on road segments using a road-network graph, a closure query, and a corresponding normal-traffic scenario. CSTP represents departures from normal traffic through a compact physical state. Queue storage captures vehicles retained on each segment, while learned estimates of congestion-wave propagation, inventory depletion, and junction interactions determine how congestion develops over time and across the network. We learn road profiles and wave behavior from normal traffic. Separate closure scenarios calibrate only two response multipliers. Vehicle counts are computed from the resulting state, while learned prediction heads estimate speed and flow. We evaluate individual mechanisms using matched simulations and conduct grouped SUMO experiments in which test locations are excluded from parameter fitting and model selection. Relative to the Cell Transmission Model (CTM), CSTP reduces the mean absolute error of local vehicle counts by 28.8% and regional count error by 9.6%, with improvements in every tested local road group. Count error increases only slightly as the number of closure labels decreases, and CSTP retains the lowest core count error across the tested label budgets. PeMS incident records provide an additional real-world evaluation of speed and flow prediction. Repeated CPU tests show that CSTP is 5.2 faster at inference and uses 49% less peak memory than CTM combined with learned prediction heads.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Conterfactual traffic prediction, Road closure effects, Physically ground machine learning, Structural interventions, Congestion propagation, Spatiotemporal modeling

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_58miOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the traffic prediciton under road-closre. Because of the limited real-world data of road-closure is rare, the authors use normal traffic data to learn basic dynamics and use small number of closed data to calibrate model. Route parameters and wave are learned from normal traffic, while closed route only calibrate two operators. Experiments are conducted on SUMO with real data.

Critical strengths
  1. The problem is meaningful.
  2. The mechanisms have clear traffic meaning. And each mechanism is validated on SUMO.
Critical weaknesses
  1. Most mehanisms are hand-crafted, including traffic wave, queue, spillback, downstream depletion, merge competition. They are meaningful in traffic, but for general ML in ICLR, these domain knowledge are hard to generalize.
  2. Model needs normal traffic scneario as reference. However, this requirement is hard to satisfy. Road-closure happens in accident, and there usually have no normal traffic records.
Questions for authors

See Weaknesses

1 comment

Thank you for your constructive comments and for recognizing the importance of the problem and the physical meaning of the mechanisms. We address your two concerns below.

W1. Machine-learning contribution and generalizability

Your concern raises the distinction between domain-specific equations and a transferable learning approach. The traffic equations are domain-specific. The broader ML contribution is how CSTP combines normal-traffic learning with an explicit intervention structure to predict closures for which few training examples are available. This matters because predicting an unseen closure requires more than extrapolating normal traffic. The model must represent how blocking a movement changes accumulation and propagates effects through connected roads. CSTP specifies these relationships through its physical state, while learning road profiles and state-dependent propagation from normal observations. Closure examples then calibrate a small shared response instead of having to teach the model the entire chain of effects. The resulting learning principle is to separate reusable system dynamics from intervention-specific calibration. This addresses a general problem in ML: learning to predict intervention outcomes when observations of normal operation are much more available than intervention examples. The matched-supervision experiment in Table 3 supports this claim by showing that the structured representation retains count accuracy at unseen closure locations with few training scenarios. We therefore see the contribution as a way to organize learning and support generalization under limited intervention data. Applying this principle elsewhere would require mechanisms appropriate to that domain, rather than directly transferring the traffic equations.

W2. Availability of the normal-traffic reference

We agree that the future traffic trajectory without an accident cannot be observed. However, the normal reference in CSTP is an estimate of that trajectory. It does not require a simultaneous no-accident record or a previous occurrence of the same closure. For example, when an accident occurs during the morning period, an expected normal trajectory can be constructed from comparable historical mornings or supplied by a normal-traffic forecasting model. The closure query then specifies the affected road and onset, and CSTP predicts the changes relative to that estimated scenario. This distinction also explains our evaluation design. Matched no-closure SUMO simulations provide a known reference to isolate closure-response accuracy. In PeMS, the reference is instead constructed from prior same-weekday speed and flow observations, with an inventory proxy derived from those profiles and road geometry. This demonstrates how the required inputs can be assembled without observing future normal traffic. Because an estimated reference can contain errors, Appendix C.1 tests timing and magnitude perturbations and finds that the local comparative advantages persist under the tested conditions. The practical requirement is therefore access to usable historical observations or a credible normal-traffic forecast. CSTP predicts the closure response conditional on that estimate.

Reviewer_tjYJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies what-if prediction of traffic under road closures. Given a road graph, a closure query, and a normal traffic scenario without closures, the proposed model CSTP predicts vehicle inventory on road segments. CSTP represents departures from normal traffic through a compact physical state. It tracks vehicle queues, how congestion spreads, decreases in vehicle counts, and interactions at junctions. The results shows that CSTP achieves competitive performance on SUMO simulated data and PeMS dataset.

Critical strengths
  1. The paper is clearly written and easy to follow. The model structure figure is clear, the express is precise and accurate, so it is easy to see what the model does.
  2. The experiments are thorough and the results are clearly visualized. The authors test each part of the model separately in controlled SUMO simulations. The main benchmark also includes cross-validation where test closure locations are never used for training. And there is complete ablation studies, robustness checks, and runtime comparison.
  3. The method needs very little closure data. For vehicle counts, only two parameters are tuned on closure cases, and the count error barely changes when the training data is cut from all groups to two. It is also fast and easy to deploy on CPU.
Critical weaknesses
  1. The contribution to machine learning is limited. CSTP is mostly a set of hand-designed traffic rules without neural networks. The learned parts are only one regression model for wave speed, two tree models for speed and flow, and two parameters chosen by grid search. Many other constants and thresholds are also set by hand. The paper does not explain what the ML community can learn from this work beyond this specific application. It also does not show whether these hand-set rules still work on a new road network for vehicle counts on a new real road network.
  2. The model makes strong assumptions about traveler behavior. It assumes everyone keeps their original trip, so nobody leaves earlier or later, cancels the trip, or switches to another travel mode. Rerouting is also handled in a very simple way. Some blocked vehicles can leave the queue by taking another turn, but the model does not add them to the roads they turn onto, so it cannot predict extra congestion on detour roads. For planned closures that are announced in advance, these behaviors can have a large effect.
  3. Real-world validation is limited. Most results come from SUMO simulation: 16 closures in one region of Nagoya, plus 10 less critical ones. The real-world PeMS dataset has only 10 incidents on 7 dates. It cannot evaluate vehicle count because counts are not available in this data. So it is still unclear whether the count improvements hold in real traffic.
Questions for authors
  1. Can CSTP model be extended to send rerouted vehicles onto detour roads? Can it also handle travelers who change departure time, cancel trips, or switch travel modes? Which parts of the model would need to change?
  2. CSTP has no neural network and very few parameters fitted on closure data, but it still beats the deep learning baselines. In my view, the closure data is small and not very diverse, so if CSTP builds traffic physics directly into the model, this will help a lot. Could the authors explain this, and show results with more closure cases or on other road networks and datasets?
  3. Has any part of the model been trained on real closure data? In the PeMS dataset, although the wave and junction rules are fitted on real normal traffic, the closure parameters and the speed/flow models still come from SUMO. Could the authors train these parts on real scenarios? Or explain the generalizability of CSTP with more experiments results.
1 comment

Thank you for your constructive comments and positive assessment of the experiments, and efficiency. We address your three questions below.

Q1. Rerouting and behavioral adaptation

Your question identifies two extensions: redistribution within the road network and changes in travel demand. CSTP currently estimates the physical closure response conditional on a supplied normal scenario. It does not predict advance changes in departure time, trip cancellation, or travel mode. To represent detour congestion, aggregate queue relief would need to be replaced by movement-level diverted flows. Vehicles leaving an upstream queue would enter the corresponding receiving-road inventory, subject to available capacity, and their effects would propagate through subsequent roads. The inventory updates and flow outputs would need consistent movement flows. Queue storage, wave propagation, and depletion would provide the starting structure, with an added route-choice component. Departure-time and mode changes require a demand-response component that shifts network entry flows over time or reduces vehicle demand. These changes could be supplied as planning scenarios or learned from behavioral observations. The resulting demand and route choices would then drive the network response. These extensions would require implementation and evaluation.

Q2. Machine-learning contribution and broader evaluation

We agree with your interpretation that physical structure is particularly helpful when closure examples are limited. This is the setting CSTP is designed to address. Task alignment also matters. The compared methods were adapted to our setting because their original input requirements or prediction targets differ. CSTP specifically targets road-level inventory prediction at unseen closure locations using a road graph and normal traffic, without requiring OD demand or route records. This alignment likely contributes to CSTP’s advantage and motivates its ML contribution. CSTP learns reusable road profiles and propagation relationships from normal traffic, represents closure effects through a compact physical state, and calibrates a small shared response from closure examples. The broader lesson is that an intervention-specific representation can support generalization when examples of the intervention are scarce. The matched-supervision experiment supports this design choice. CSTP retains nearly its full-budget count accuracy with two training groups and has the lowest core count error at every tested budget (Table 3). Shared rules and thresholds remain fixed across test closures, while road geometry and normal observations provide location-specific profiles. These results support the value of the representation under the shared input and supervision constraints. The additional non-critical cases also test a different congestion regime. In Table 21, CSTP has lower local count MAE than both CTM and IGSTGNN in all ten held-out road groups. These cases broaden the evaluation within Nagoya. PeMS further tests transfer on a different network using the available speed and flow measurements. We therefore distinguish cross-location count generalization from real-network speed and flow transfer.

Q3. Training on real incidents and generalizability

No CSTP component is fitted to real incident outcomes in the reported PeMS evaluation. Road profiles, wave estimates, and junction calibration use prior normal traffic, while the closure-response parameters and speed/flow heads remain fitted on SUMO. This setup deliberately tests whether the learned closure response transfers without incident-specific fitting. Under this setup, CSTP achieves the lowest mean local speed and flow MAE and RMSE among the compared methods (Table 2 and Appendix F). This provides evidence that the transferred response remains useful with real observations and estimated normal references. It does not directly validate vehicle-count accuracy in real traffic. Real incident labels could be used to fine-tune the speed/flow heads. The physical response could also be calibrated through speed/flow prediction loss, although this would differ from the current count-based calibration. Such an adaptation study should separate training and test incident dates because same-date records can share observations. It would assess speed and flow adaptation, while direct validation of inventory would still require observed road-level counts.

Reviewer_pRE2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CSTP, a structured traffic prediction framework for estimating road-closure effects from a road-network graph, a closure intervention, and a supplied normal-traffic reference. CSTP represents closure-induced deviations through physical mechanisms involving queue storage, congestion-wave propagation, inventory depletion, and junction interactions. The resulting state directly determines vehicle counts, while separate learned heads predict speed and flow.

Critical strengths

S1. This paper focuses on predicting traffic responses to road closures without requiring historical observations from the target closure. The proposed formulation addresses an important practical challenge in traffic disruption planning. Incorporating explicit traffic mechanisms also provides useful inductive biases compared with purely data-driven forecasting architectures.

S2. This paper includes controlled mechanism tests, comparisons with physical and neural baselines, grouped evaluation across closure locations, component ablations, and computational efficiency analyses. The improvements in vehicle-count prediction are substantial, particularly near closures.

Critical weaknesses

W1. CSTP requires a complete normal-traffic reference over the forecasting horizon, including road-level vehicle counts, speeds, and flows. In the matched SUMO experiments, these reference trajectories are obtained from corresponding no-closure simulations under the same background conditions. However, in practical road-closure planning, the future normal-traffic trajectory is itself unobserved.

W2. A central contribution is accurate forecasting with limited closure data. However, the two-group setting in Appendix C.2 still uses 96 temporal observations per closure case across the core road network. Therefore, even a small number of closure groups can provide a substantial number of supervised road-time observations. Furthermore, the physical profiles and wave estimates are learned from a large collection of normal-traffic observations. These results support the claim that CSTP requires relatively few distinct closure scenarios, but they do not necessarily demonstrate data efficiency.

W3. in CSTP, the final state updates and prediction heads do not enforce exact network-wide vehicle conservation. Moreover, Appendix A indicates that alternative-movement relief is represented in aggregate form without explicitly propagating diverted inflows to their receiving roads. This limitation is important for road-closure planning because closures can redistribute demand and create congestion on alternative routes. it remains unclear whether the proposed approximation can reliably characterize more complex traffic responses involving substantial rerouting.

W4. The ablation results indicate that queue storage is responsible for the majority of the improvements in vehicle-count prediction. In Table 5, removing queue storage increases local count MAE from 10.16 to 28.55, whereas removing the learned wave law, inventory depletion, or junction interactions produces considerably smaller degradations. THese raise questions about the incremental benefit of the more elaborate dynamics beyond a relatively simple queue-storage model.

Questions for authors

NA

1 comment

Thank you for your constructive comments. We provide our point-by-point responses below.

W1. We agree that future normal traffic is unobserved. CSTP predicts closure-induced changes conditional on an estimated no-closure scenario, which can be constructed from historical profiles or a normal-traffic forecasting model. The matched SUMO simulations provide a controlled setting for isolating closure-response accuracy. To examine prediction with an estimated reference, our PeMS evaluation uses speed and flow profiles from prior same-weekday observations. Because PeMS does not provide road-level inventory, we approximate it from flow, speed, and road length, subject to storage bounds. This supplies the inventory input needed to represent accumulation and storage use without observing future normal traffic. Using this estimated reference, CSTP achieves the lowest mean local speed and flow errors among the compared methods, with its SUMO-fitted response parameters and prediction heads held fixed (Table 2). Appendix C.1 further examines reference uncertainty and shows that local comparative advantages persist under the tested timing and magnitude perturbations. Together, these evaluations support the use of estimated references under the tested conditions.

W2. We appreciate your distinction between the number of closure scenarios and the number of observations. Our intended claim concerns accurate prediction at unseen closure locations using few distinct training scenarios. Each scenario still contains many road-time observations, and normal-traffic learning also requires data. This distinction reflects the different roles of the two data sources. Normal observations supply reusable road profiles and wave estimates, while closure observations calibrate two response multipliers for the count model. This lets CSTP reuse normal-traffic learning when predicting a closure at a different location. We test this benefit under matched supervision in Appendix C.2. Every method receives identical sampled road-time observations at each scenario budget, and each test group is excluded from fitting and model selection. With only two training groups, CSTP retains nearly its full-budget count accuracy and achieves lower core count error than the compared controls using their full training budget (Table 3). The result supports reduced dependence on distinct closure scenarios after normal-traffic learning. It does not imply that the total observation count is small.

W3. CSTP estimates aggregate closure effects without requiring OD demand or route records. Its updates represent queue accumulation, spillback, downstream depletion, and junction interactions, with road-storage bounds. These road-level constraints do not enforce exact network-wide conservation. In particular, alternative-movement relief estimates reduced blockage and upstream queue release without explicitly assigning diverted inflows to receiving roads. The separate flow head is also not constrained to balance the inventory updates. Our experiments assess the predictive value of this aggregate approximation. The grouped benchmark supports accurate inventory prediction across the evaluated closure locations, while controlled tests isolate the modeled upstream, downstream, and junction responses. These findings support rapid assessment of aggregate closure impacts with limited inputs. They do not establish reliability when substantial detour traffic dominates congestion, which requires explicit receiving-road flows coupled to inventory updates.

W4. To directly address your question about the benefit beyond queue storage, we conducted an additional comparison using the original four road-group folds. The queue-storage baseline retains finite storage, storage-triggered spillback, and aggregate alternative-movement relief, while removing traversal delay, inventory depletion, and junction interactions. Its response parameter was calibrated independently using only each fold’s training groups. Against this calibrated baseline, the unchanged submitted CSTP model achieves 9.4% lower local count MAE. To make the comparison more demanding, we also tested a baseline with separately calibrated constant-speed propagation. CSTP retains a 7.0% reduction in local count MAE. Both comparisons also favor CSTP in core count MAE and in core and local count RMSE. These comparisons directly measure the additional mechanisms’ joint contribution, complementing the individual removals in Table 5. Queue storage provides the foundation for accumulation, while propagation, depletion, and junction interactions improve predictions of how closure effects develop across connected roads.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.