acceptodds
Under review as a conference paper at ICLR 2027

When Events Reshape Traffic Graphs: Event-Triggered Dynamic Graph Learning for Traffic Forecasting

Abstract

With the increasing complexity of urban traffic systems and the growing availability of spatio-temporal observations, accurate traffic forecasting under dynamic and event-driven conditions has become a critical challenge. Most existing dynamic graph forecasting methods update spatial dependencies according to fixed time intervals or historical traffic correlations, allowing them to capture that traffic relationships have changed but making it difficult to identify whether such changes are triggered by accidents, road construction, severe weather, or large-scale events. Moreover, event impacts are rarely isolated to individual locations, but instead propagate through upstream and downstream road segments and may further induce traffic redistribution over alternative routes. Existing methods also tend to model long-term traffic regularities and short-term event disturbances within a unified representation, causing transient anomalies to interfere with stable trends and periodic patterns. To overcome these limitations, we propose ETDGN, an Event-Triggered Disentangled Graph Network for traffic forecasting. ETDGN introduces an event-triggering mechanism that jointly exploits explicit contextual information and implicit traffic anomalies to selectively activate graph updates when event-related changes occur. An event-conditioned dynamic graph is then constructed to capture event-induced spatial dependency shifts and model the propagation of disturbances over the traffic network. To prevent transient events from contaminating stable traffic dynamics, ETDGN further disentangles long-term trend and periodic components from short-term event residuals, and incorporates event-aware residual correction into the routine forecast. Extensive experiments demonstrate that ETDGN outperforms state-of-the-art baselines in both forecasting accuracy and scalability, validating its effectiveness for continual spatio-temporal forecasting. Code is available at https://anonymous.4open.science/r/ETDGN-ICLR-2027-C367.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Traffic Forecasting, Event-Triggered Dynamic Graph Learning, Spatio-Temporal Graph Neural Networks, Event Impact Propagation

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

This paper proposes ETDGN, an Event-Triggered Disentangled Graph Network for traffic forecasting. The main idea is to move dynamic graph adaptation from a purely time-driven or correlation-driven mechanism to an event-triggered mechanism. ETDGN first builds a routine traffic reference by decomposing traffic observations into trend, seasonal, and event-residual components. It then estimates event evidence from explicit context, traffic residuals, and spatially coherent abnormalities. When event evidence is sufficiently strong, the model constructs an event-conditioned graph, propagates event influence over this graph, and uses a gated residual branch to correct the routine forecast.

Overall, the paper addresses an important and practical problem in traffic forecasting: spatial dependencies may change abruptly due to accidents, weather, holidays, or other external events. The proposed framework is coherent, well motivated, and supported by extensive experiments.

Critical strengths
  1. The motivation is clear and compelling. The paper convincingly argues that dynamic graph updates should not be driven only by time progression or historical correlations, since real traffic dependencies are often reshaped by external or latent events.

  2. The model design is coherent and well aligned with the stated challenges. The event trigger, event-conditioned graph, propagation module, memory retrieval, and gated residual correction each serve a clear role in the overall architecture.

  3. The separation between routine dynamics and event-induced deviations is a strong design choice. Modeling event effects as residual corrections over a routine forecast is intuitive and helps avoid contaminating stable trend and periodic patterns.

  4. The experimental results are strong across multiple datasets, metrics, and forecasting horizons. The ablation study also supports the usefulness of event-triggered graph updating, event-conditioned graph construction, propagation, and disentanglement.

  5. The qualitative analyses are useful. The event-pattern case study and propagation visualizations provide intuition about how the model responds to event-related conditions and how event influence evolves over space and time.

Critical weaknesses
  1. For datasets without explicit event annotations, the notion of “event” could be more clearly grounded. The paper appears to derive event supervision from routine forecasting residuals in such cases. This is a reasonable practical strategy, but it makes the event trigger closer to a residual-abnormality trigger than a directly verified external-event trigger. The paper would benefit from a clearer discussion of this distinction.

  2. The ablation study shows that the event branch and propagation module improve forecasting accuracy, but it does not report how often the event branch is activated. Since the central claim is event-triggered computation, activation statistics would be important for interpretation. I suggest reporting the event-branch activation rate on each dataset, ideally broken down by normal vs. event-related periods when external contexts are available. It would also be useful to correlate trigger intensity or activation rate with forecasting improvement, to show whether the event branch mainly helps during event-affected intervals rather than simply acting as additional model capacity.

  3. The qualitative propagation analysis is compelling, but stronger quantitative validation would further support the interpretability claim. For example, if real incident locations, weather-affected regions, or event windows are available, it would be helpful to evaluate whether the learned event propagation aligns with these external signals.

Questions for authors
  1. What are the event-branch activation rates on PEMS-Stream, CA-Stream, PEMS08, and TaxiBJ?
  2. On datasets without explicit event annotations, how sensitive is the method to the residual-based pseudo-event threshold?
  3. Can the authors report activation rates separately for normal periods and event-related periods on TaxiBJ?
  4. Does the event branch contribute most of the performance improvement during high-trigger-intensity intervals?
  5. Can the learned event propagation be quantitatively compared with real incident, weather, or holiday-event signals when such information is available?
Reviewer_uDUnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

External events, such as traffic accidents, road construction, and extreme weather, can alter local traffic conditions and affect how traffic disturbances propagate through road networks, ultimately changing the spatial dependencies among traffic nodes. However, existing methods often overlook the influence of these external factors. Motivated by this limitation, this paper proposes ETDGN to model event-induced changes in traffic dynamics and spatial dependencies.

Critical strengths
  1. The paper addresses a practically relevant problem in traffic forecasting, as external events such as accidents, road construction, and extreme weather can disrupt regular traffic patterns and alter spatial dependencies. Considering event-related disturbances when modeling dynamic traffic relationships is meaningful and potentially useful for real-world traffic forecasting.

  2. The experimental results demonstrate that ETDGN achieves competitive forecasting performance across three traffic datasets, with consistent improvements over the compared baselines in terms of MAE, RMSE, and MAPE. The paper also provides ablation studies and efficiency evaluations, which offer some empirical support for the effectiveness and computational efficiency of the proposed framework.

Critical weaknesses
  1. The methodological novelty of ETDGN appears somewhat limited, as several of its core ideas have already been explored in previous studies. For example, ConFormer (KDD 2026) [1] investigates accident-informed traffic forecasting and graph-based impact propagation, while HyperD (AAAI 2026) [2] focuses on separating regular periodic patterns from irregular traffic disturbances. Furthermore, the paper does not clearly distinguish its core contribution of selective event-triggered graph updating from existing approaches, making the proposed method appear more like a combination of established techniques.

  2. Although ETDGN claims to model traffic events, based on the formulations and descriptions in the paper, the proposed mechanism seems to rely more heavily on prediction residuals. Specifically, in Appendix D.3, pseudo-event labels are generated by thresholding routine prediction errors, which may not necessarily correspond to actual traffic events. Without real event annotations or independent validation, it remains unclear how the proposed event-triggering mechanism fundamentally differs from residual-triggered graph adaptation. Furthermore, the graph propagation mechanism in Eq. (9) does not sufficiently substantiate the claimed temporal delays, directional effects, or recovery dynamics. Similarly, the residual correction in Eq. (11) does not necessarily ensure that event-related information is effectively separated from routine representations.

[1]Wang H, Yong J, Wang J, et al. Towards Resilient Transportation: A Conditional Transformer for Accident-Informed Traffic Forecasting[C]//Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2026: 1424-1434.

[2]Shao M, Zhang Z, Wang Y, et al. Hyperd: Hybrid periodicity decoupling framework for traffic forecasting[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2026, 40(18): 15689-15697.

Questions for authors

Q1. Could the authors clarify the key methodological differences between ETDGN and existing event-aware traffic forecasting methods, particularly ConFormer (KDD 2026) and HyperD (AAAI 2026)? Beyond combining event-conditioned graph learning with routine–event disentanglement, what is the main technical novelty of ETDGN?

Q2. To my knowledge, PEMS-Stream, CA-Stream, and PEMS08 do not provide ground-truth event annotations. Does this mean that the event labels for all three datasets are generated from prediction residuals? If so, how do the authors distinguish actual traffic events from ordinary prediction errors or measurement noise? The authors could consider validating the proposed event-triggering mechanism on a dataset containing real traffic incident records, such as TraffiDent [1] (NeurIPS 2025). I would also appreciate clarification on how the proposed event-triggered graph adaptation fundamentally differs from residual-triggered graph adaptation.

Q3. Eq. (9), introduced around line 263, defines the event propagation mechanism, which essentially resembles a graph propagation operation with residual connections. While graph propagation can certainly enable multi-hop information diffusion, it does not necessarily demonstrate that the model accurately captures the actual direction of event propagation or the temporal delays between affected road segments. In particular, Figure 1(c) illustrates event propagation with delays of approximately 1, 2, and 3 hours. However, the (L)-step graph propagation in the proposed method does not directly correspond to (L) hours or any specific temporal interval. In my view, expanding the spatial receptive field is not necessarily equivalent to modeling temporal propagation delays. Could the authors clarify how the model identifies and captures such time-lagged propagation effects through spatial message passing?

Q4. The formulation around line 294 expresses the final prediction as the sum of a routine forecast and an event-aware residual correction. The paper interprets this design as a way to prevent short-term event disturbances from contaminating long-term trends and periodic patterns. However, adopting a residual forecasting structure does not necessarily imply that the routine branch is free from event-related information. In fact, the initial representation (H_t^0), defined around line 172, incorporates the original input, trend, seasonal component, and event residuals. The routine graph learner then constructs (A_t^r) based on this representation. Therefore, even though the final prediction uses two separate branches, event-related information may still enter the routine branch through its upstream inputs. Could the authors further clarify how the proposed method ensures effective separation between routine traffic dynamics and event-related disturbances?

[1]Gou X, Li Z, Lan T, et al. TraffiDent: A dataset for understanding the interplay between traffic dynamics and incidents[J]. Advances in Neural Information Processing Systems, 2026, 38.

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

The paper addresses a meaningful problem: traffic disturbances should not necessarily be handled in the same way as ordinary temporal variation. Its most promising contribution is selectively activating an event-specific graph and residual forecasting branch using contextual information, traffic abnormalities, and persistent event states.

However, the methodological novelty is narrower than the paper’s framing suggests. Dynamic graph learning, traffic decomposition, event-aware modeling, graph diffusion, and gated prediction are already established directions. The potential originality lies in their particular integration into a conditionally executed architecture—not in introducing event-aware traffic forecasting or disentangled dynamic graph learning generally.

My main concern is therefore not the experimental coverage, but the gap between the claimed conceptual advance and what the mathematical formulation establishes.

Critical strengths
  1. The paper raises a valuable question: under what conditions should anomalous observations trigger additional graph structure adaptation, rather than merely generating different predictions on the original graph?

  2. The combination is coherent.

  3. Algorithm 1 explicitly skips event-graph construction, event propagation, and memory retrieval during inactive periods. This provides a credible contribution beyond continuously reweighting an always-active model.

Critical weaknesses
  1. For datasets without event annotations, Appendix D.3 generates event pseudo-labels by thresholding the residuals of routine estimation. The trigger is therefore effectively trained to identify whether the current observation deviates substantially from the learned routine reference. This supports anomaly-sensitive graph adaptation, but cannot determine whether such deviations stem from traffic accidents, severe weather, or sensor noise. Even when external event context is available, conditioning the graph structure on such context does not automatically imply that the causal effect of events on spatial dependencies has been identified. I do not believe that traffic forecasting methods must adopt a causal framework.

  2. While the most relevant work is cited, the paper does not sufficiently articulate the substantive differences. For instance, ConFormer, which combines accident information with graph propagation and conditional modeling, is cited — yet the authors should clarify what event-to-graph-structure mechanism is introduced here beyond selective activation and the support for implicit event evidence.

Questions for authors
  1. What is the relationship between graph propagation depth, actual elapsed time, and upstream–downstream propagation delays?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.