acceptodds
Under review as a conference paper at ICLR 2027

RIDE: An Open Dataset and Benchmark for Train Delay Prediction

Abstract

Train delay prediction is an important problem for both passengers and railway operators, yet progress in the field remains difficult to assess due to the lack of standardized datasets, prediction targets, and evaluation protocols. To address this gap, we introduce RIDE, an open dataset and benchmark for train delay prediction built at nationwide scale over the Belgian railway network. RIDE covers 94.5M train events, 3.6M journeys, and 35.7M weather records from 2023 to 2025. It is organized as a layered data pipeline from raw railway and weather sources to two public releases: a reusable intermediate relational dataset and model-ready benchmark datasets. The benchmark standardizes the prediction task and the training and testing data. It also provides a unified evaluation protocol that supports direct comparison across models. Using this framework, we provide the first comprehensive evaluation of representative non-learning, tree-based, and deep learning model families under a shared setup. The resulting comparison reveals distinct behavior across model families: learning-based models clearly outperform non-learning approaches, graph neural networks achieve the best mean performance, yet the strongest sequential, attention-based, and graph-based models remain surprisingly close overall. Beyond aggregate mean absolute error (MAE) and root mean squared error (RMSE), the framework also provides breakdowns by prediction horizon and delay change, enabling more detailed analysis of model behavior across forecasting regimes.

Primary area:
datasets and benchmarks
Keywords:
Datasets, Benchmark Evaluations, Train Delay Prediction, Railway Transportation, Spatiotemporal Forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gJWGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces an open dataset, along with a few benchmarks, for models, metrics, and problem formulation. The authors provide a detailed explanation of their tier-based pipeline, as well as some preliminary experimentation with (mostly vanilla) models. Hyperparameters are tuned with optuna (which method? BO?), and the results show that XGboost and non-learning baselines have the best performance in short-term/first-order delay propagation prediction, while GNNs are capable in more nuanced scenarios.

Critical strengths

This is an excellent dataset, and I agree that comparability and gold-standard datasets are absolutely fundamental to true research progress. This problem is still a niche, but I don’t think that serves as an argument to minimise this effort.

The authors try to make this a very thorough paper, covering weather, delays, and a 4-layered pipeline. I really liked the model-specific datasets (tabular, sequential, GNN, Graph-event), which help structure the field into distinct task types and methods. The results generally make sense and are coherent with the state of the art. The hyperparameter tuning is well argued, and the results are reasonably well discussed.

The paper also provides an extremely detailed (and long!) appendix with all details.

Critical weaknesses
  • Although I’d still count it as a contribution, the dataset mentioned is already open, so the “value added” of the paper alone is questionable as a “dataset and benchmarks” paper.
  • d^{last}{s,i} definition is confusing. First, d^{last}{s,i} is different from d_{s, i, n}, correct (n is also a last event prediction…)? t^s is the observation that maps to t^{sched}{s,i,1}, right (if they are equal, delay is zero; I mean you say that “it is the scheduled time of the next event”, which next event?)? In which circumstance can t^s-t^{sched}{s,i,1} be different from d_{s,i,−1}?
  • The baseline models are quite vanilla. Did the authors check which of the existing models mentioned in the literature already provide code? If they are currently the state of the art, why not include them as well?
Questions for authors

Look at the weakness section.

4 comments

Thank you for your review. We appreciate your emphasis on the importance of comparable datasets for research progress, and your positive assessment of RIDE’s model-specific datasets and detailed appendix.

W1

While we agree that the underlying railway and weather records are open, preparing them for train delay prediction is substantial work and a central contribution of RIDE. This includes cleaning and aligning train events, reconstructing the network topology, and integrating weather data. RIDE performs this work at nationwide scale and releases both a reusable relational dataset and model-ready datasets. Researchers can build on these releases instead of repeating the data preparation themselves. We believe that saving researchers this substantial effort constitutes significant added value.

W2

Yes, ds,i,nd_{s,i,n} is the delay at the nn-th future event, whereas ds,ilastd_{s,i}^{\mathrm{last}} is the last known delay, explained below. tst^s is the time at which the prediction is made for the snapshot ss. For each train ii, we distinguish events that have already been observed from events that have not yet been observed. The previous observed event has index −1-1, and the first future planned event has index 11. Thus, ds,i,−1d_{s,i,-1} is the observed delay at the previous event, whereas ts,i,1schedt_{s,i,1}^{\mathrm{sched}} is the scheduled time of the next event in the train's itinerary. With the same indexing, ts,i,−1schedt_{s,i,-1}^{\mathrm{sched}} would be the scheduled time of the previous observed event. To clarify how these quantities relate, consider how ds,ilastd_{s,i}^{\mathrm{last}} is computed. If the train kept its delay from the previous observed event, the next event would occur at ts,i,1sched+ds,i,−1t_{s,i,1}^{\mathrm{sched}}+d_{s,i,-1}. Subtracting this implied time from tst^s tells us whether it lies before the snapshot. If the implied time falls before the snapshot, it conflicts with the fact that the event has not yet been observed. We therefore increase the carried-forward delay by that gap, placing the implied event time at tst^s. This gives a more informative estimate of the current delay of the train: an event that is still unobserved after its scheduled time must be at least that late. This is expressed by ds,ilast=ds,i,−1+max⁡(0,,ts−ts,i,1sched−ds,i,−1)d_{s,i}^{\mathrm{last}}=d_{s,i,-1}+\max(0,\\,t^s-t_{s,i,1}^{\mathrm{sched}}-d_{s,i,-1}), which can be written more intuitively as:

\left\\{\begin{array}{ll} d_{s,i,-1} & \mbox{if }t_{s,i,1}^{\mathrm{sched}}+d_{s,i,-1}\geq t^s,\\\\ d_{s,i,-1}+\bigl[t^s-(t_{s,i,1}^{\mathrm{sched}}+d_{s,i,-1})\bigr] & \mbox{else.} \end{array}\right.$$ We will include these clarifications in the paper, and revise notation where needed. **W3** We checked whether public code was available for the prior methods discussed in the paper. The public implementations we found that fit our benchmark are already used in our MLP and Transformer models [1]. We could not find public implementations for the other cited methods. Additionally, published results use different datasets and prediction tasks, so they do not establish a clear state of the art. Our implementations represent the main model families used in the literature and provide a common starting point for evaluating future methods. **Reference** [1] Elliker et al. (2026), [Simulation-Driven Railway Delay Prediction: An Imitation Learning Approach](https://ojs.aaai.org/index.php/AAAI/article/download/39239/43200).

Dear authors,

Thank you so much for answering my comments.

Regarding W2, it is now very clear. Please make it clear in the paper (perhaps with a diagram).

W1. Yes, of course, providing data treatment for upcoming research is a contribution. And I didn't say either that this paper wouldn't contribute at all to the "datasets and benchmarks", but my point still stands that, for ICLR, it is not as strong an argument as if you were providing a new dataset W2. This also relates to W1. The "benchmark" argument for ICLR would be much stronger with stronger baselines.

I am still leaning toward keeping the score and am quite willing to hear further from the authors to convince me otherwise. Which other baselines could be added? Even though there are no public implementations, is there enough information in any such paper to build your own version?

Thank you for the follow-up. We will make the requested W2 clarifications in the paper. We understand the distinction you draw between this work and collecting a new source dataset. Regarding the benchmark, we believe our baselines are representative of the main approaches in train-delay research and are more substantial than their short model-family names suggest. Published methods provide enough detail to guide RIDE-specific implementations, as we did for the Graph-event, LSTM, and GNN baselines. Their original architectures cannot be added unchanged because the settings for which they were designed differ from RIDE’s. Below we explain how each baseline was selected and implemented.

Translation: Included as a simple reference method that carries the known delay forward, following similar baselines in [1, 2].

Graph-event: We draw on event-based delay propagation models [3, 4], but their simulation rules depend on the topology, event definitions, and operational information of their datasets. Those rules cannot be applied directly to RIDE, so we had to design and implement our own Graph-event model using RIDE’s topology and train-event records.

XGBoost and MLP: We retained the implementations from [5]. Both are standard methods whose existing prediction setups could be used for RIDE, with routine tuning rather than a new architecture.

LSTM: Prior LSTM studies show how train operations can be modeled as sequences [6–8], but their architectures reflect their particular inputs and prediction tasks. The closest is Sobrie et al. [6], who use an encoder–decoder across the Belgian network to predict five future signal passings from three observed passings. In RIDE, we predict the next 15 scheduled events of every active train from a sampled network snapshot. We extend their encoder–decoder design by giving each decoder step the known features of its future event and using attention pooling over the observed history, rather than repeating one encoded summary across steps. The input also requires change: on the individual lines studied in [7,8], trains ahead of a target form a natural sequence. Across RIDE’s full network, relevant trains can join from different routes along the target’s itinerary, so we summarize surrounding traffic as local features encoded by a separate static-context MLP. Sobrie et al. also use train and spatial features, but feed them into the past-signal sequence rather than a separate context branch. Our resulting LSTM builds on prior sequence models while adapting their inputs and decoder to RIDE’s snapshot setting and multi-event target.

Transformer: We retained the architecture introduced in [1] and used the implementation released with [5]. Its train-token design already fits RIDE’s task of predicting future event delays for multiple trains.

GNN: Prior GNNs show how graph structure can capture interactions in railway operations [9–11], but their architectures reflect the tasks and inputs they were designed for. SAGE-Het and NRI-GraphSAGE predict train delays after fixed time intervals, while Huang et al. represent train events as nodes to model delay propagation. In RIDE, we must predict delays at the next scheduled events of every active train, so we represent those future events as train-to-station edges and make predictions on them. The available inputs also shape the graph. For example, inferred routes let us estimate where trains are on the railway graph at prediction time and derive features for the links they traverse. Using these and other available features requires choices about graph structure, feature placement, and message passing, so we could not use the published architectures unchanged. We also tested GAT and GraphSAGE layers, which are used in the cited papers, on our graph. GINE performed best in these tests, so we retained it [12]. The resulting GNN is a substantial baseline grounded in prior train-delay models and designed for RIDE’s inputs and prediction task.

The labels LSTM, Transformer, and GNN identify broad model families, but do not convey the specific architectures evaluated here. The Transformer follows a published train-delay design, while we designed the encoder–decoder LSTM and heterogeneous GNN for RIDE’s inputs and prediction task, drawing on recent train-delay research. We will describe these designs more explicitly in Section 4.4 so their names do not suggest generic implementations.

References

[1] Arthaud et al., Transformers à Grande Vitesse: Massively Parallel Real-Time Predictions of Train Delay Propagation, 2024.

[2] Wang and Work, Data Driven Approaches for Passenger Train Delay Estimation, 2015.

[3] Goverde, A Delay Propagation Algorithm for Large-Scale Railway Traffic Networks, 2010.

[4] Wei et al., Modeling Cascade Dynamics of Railway Networks under Inclement Weather, 2015.

[5] Elliker et al., Simulation-Driven Railway Delay Prediction: An Imitation Learning Approach, 2026.

[6] Sobrie et al., Capturing Complexity over Space and Time via Deep Learning: An Application to Real-Time Delay Prediction in Railways, 2023.

[7] Gao et al., A Novel Deep Learning Model for Short-Term Train Delay Prediction, 2023.

[8] Huang et al., Modeling Train Operation as Sequences: A Study of Delay Prediction with Operation and Weather Data, 2020.

[9] Li et al., Railway Network Delay Evolution: A Heterogeneous Graph Neural Network Approach, 2024.

[10] Huang et al., Explainable Train Delay Propagation: A Graph Attention Network Approach, 2024.

[11] Li et al., Integrating Graph and Neural Relational Inference for Network-Wide Train Delay Prediction: An Equilibrium between Accuracy and Interpretability, 2026.

[12] Hu et al., Strategies for Pre-Training Graph Neural Networks, 2019.

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

The authors propose RIDE, an open dataset and benchmark for train delay prediction built at nationwide scale over the Belgian railway network, which provides a standardized data-processing pipeline and a benchmark protocol. In their experiments, the authors compare a range of methods, from non-learning approaches to learning-based models, under a shared setup and analyze the performance differences across models.

Critical strengths
  • S1: The authors construct a large-scale real-world open dataset.
  • S2: The authors build a unified benchmark framework for comparing a wide range of existing methods and analyze the performance differences among them.
  • S3: The paper is well written and easy to follow.
Critical weaknesses
  • W1: The technical novelty is limited. The problem setting is not new, and the paper is largely limited to benchmarking existing methods. The value of the work would be increased if the authors proposed a new method that addresses the limitations of existing methods revealed by the unified comparison, or at least provided design guidelines or model selection criteria for such methods.
  • W2: The benchmark data are limited to Belgium. Since delay patterns and frequencies are expected to differ substantially across countries, a comparison based on data from a single country is unlikely to provide a general performance benchmark across methods. The value of the dataset would be further increased if it allowed data from multiple countries to be compared in a unified manner.
Questions for authors
  • Q1: For predicting future train delays, it seems important to take into account the causes of delays that have already occurred (e.g., equipment failures or accidents). Is it possible to make use of such information?
  • Q2: As shown in Tables 4 and 5, even among learning-based methods, no single method consistently achieves the best performance. Based on the findings from the unified benchmark in this paper, what criteria should practitioners use when selecting a model?
  • Q3: Related to W2, would it be possible to process existing train delay datasets, such as those for China, Italy, Germany, and Finland, with the RIDE pipeline and incorporate them as part of the dataset?
3 comments

Thank you for your review. We appreciate that you recognized RIDE's large-scale open dataset, unified comparison of methods, and clear presentation.

W1(A)

Regarding the concern that the paper has limited technical novelty and largely benchmarks existing methods, we emphasize that RIDE’s main contribution is a public dataset and shared evaluation framework that make train-delay methods comparable. Three of the baseline methods were also designed and implemented specifically for RIDE. Published studies use different datasets, prediction targets, and evaluation protocols, so their scores cannot establish a clear state of the art. RIDE is the first to combine processed train-delay data with a fixed prediction task, temporal splits, a shared evaluation protocol, and baseline implementations. For most methods we considered, we could not find public code. We therefore developed the simulation-based Graph-event model, the encoder-decoder LSTM, and the heterogeneous GNN with GINE-style message passing. They draw on model families in prior work, but their specific designs and implementations are ours. Together, these releases provide reproducible reference results for future work. We will make this purpose clearer in the paper.

W1(B)/Q2

We agree that providing design guidelines and model-selection criteria is very important. Our breakdowns by prediction horizon (Table 4) and delay change (Table 5) offer guidance for practitioners with different needs. For example, updating passengers about an approaching train’s arrival calls for short-term predictions. For horizons up to 15 minutes, the LSTM is a strong choice: it leads from 5 to 15 minutes and is nearly tied with XGBoost in the first five minutes. Real-time traffic management may place particular weight on anticipating large delay increases because these can substantially disrupt planned train operations. The GNN performs best in most such cases, with the Transformer similarly strong for the most extreme increases. We will include these examples and explain the reasoning behind the model-selection guidance in our discussion of Tables 4 and 5.

W2/Q3

We recognize that model rankings from the Belgian network may not generalize to other countries. In principle, the RIDE pipeline and evaluation protocol could be extended to datasets from other countries, and we agree that this would strengthen the benchmark, but adding other countries would be a major extension that we consider a direction for future work. Germany and Finland are the most promising among the cited datasets because they span several years. However, neither release provides full network topology comparable to RIDE’s, so we would need to obtain it and align it with their train events. We would then need to adapt the processing to each country’s sources, produce event records in a comparable format, and check their quality and coverage. Together, these steps would require substantial additional work for each new network. We will clarify that further in the Limitations section.

Q1

Knowing the causes of earlier delays, such as equipment failures or accidents, could improve predictions by providing additional input features. However, our data sources provide scheduled and observed train events but do not record causes for individual trains. Research on train-delay prediction often focuses on secondary delays as they propagate through the network. Such delays can be modeled from observed train events without explicit cause labels.

Reviewer_SUBJOpenReview ↗

Official Comment by Reviewer SUBJ

Thank you for the response. It addresses most of my concerns and questions.

I acknowledge the contribution of a unified benchmark on RIDE. However, I am not convinced that the benchmark should be confined to this single dataset. If RIDE addresses an existing prediction problem rather than a new one, the fragmentation that motivates the paper calls for discussing a unified comparison that also covers the existing datasets. I understand that this would require substantial effort. Without it, however, the comparability that RIDE provides holds only on RIDE, and results on the existing datasets remain comparable neither with each other nor with results on RIDE.

For this reason, I keep my current rating.

Thank you for your follow-up.

We agree that extending standardized evaluation across multiple railway datasets would be valuable, particularly for studying cross-network generalization. However, we disagree that this is necessary for RIDE to meaningfully address the fragmentation of train delay prediction research. In fact, the existence of an established body of work using incompatible datasets, prediction targets, and evaluation protocols is precisely what motivates the introduction of a common benchmark. RIDE provides such a benchmark, enabling existing and future methods to be evaluated under the same conditions, independently of the datasets used in their original publications.

We would also like to clarify that incorporating existing datasets would not, by itself, make previously published results directly comparable. Differences in prediction targets, data splits, and evaluation procedures would still require methods to be re-evaluated under a common protocol. Extending RIDE to other countries would therefore broaden the scope of standardized evaluation and enable cross-network generalization studies, rather than resolve a limitation of the comparisons already provided by RIDE.

We acknowledge that RIDE does not standardize all existing railway datasets, but do not believe that addressing fragmentation requires resolving it entirely within a single contribution. As already discussed, extending the benchmark across countries would require substantial additional work. We therefore consider RIDE a meaningful contribution in its own right and a foundation for broader multi-dataset evaluation.

Reviewer_8aivOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RIDE, an open dataset and unified benchmark for train delay prediction using Belgian railway data from 2023 to 2025, including 94.5 million train events, 3.6 million journeys, and 35.7 million weather records. The authors provide reusable intermediate data and model-ready benchmark datasets, together with standardized prediction tasks, temporal splits, and evaluation metrics. Seven models are evaluated under a common protocol, with additional analysis across prediction horizons and delay-change regimes.

Critical strengths
  1. RIDE integrates railway operational, infrastructure, and weather data at a nationwide scale. Its layered organization provides reusable intermediate data and model-ready datasets, reducing the effort required for data preparation and supporting alternative modeling approaches.

  2. The paper establishes common prediction targets, temporal splits, and evaluation metrics for seven representative models. The additional analysis by prediction horizon and delay change provides useful insights beyond aggregate prediction accuracy.

Critical weaknesses
  1. Section 3.2 and Appendix B.4 describe the route inference procedure. Although the appendix reports a sensitivity analysis showing that inferred routes are relatively stable under different parameter settings, this does not establish their accuracy against actual train routes. Furthermore, inferred routes are used to construct downstream model features, but the paper does not directly assess how potential route inference errors affect prediction performance. This leaves uncertainty about the reliability of route-dependent data and its influence on benchmark results.

  2. The paper discusses railway datasets from China, Italy, Germany, and Finland and identifies differences in their coverage and characteristics. However, the comparison remains largely qualitative. Although RIDE contributes a reusable data resource and unified evaluation protocol, the paper could more clearly establish which specific data characteristics or research capabilities distinguish it from existing resources. A more systematic comparison would help clarify its distinctive contribution.

Questions for authors

1: Could the authors provide independent evidence of route inference accuracy, or analyze how plausible inference errors affect downstream prediction results?

2: Could the authors clarify which specific research tasks or evaluation capabilities RIDE supports that are not readily supported by existing public railway datasets?

  1. How consistent are the relative model rankings across different time periods or train categories within the test set? Could the authors clarify whether the aggregate results reflect stable performance across different operating conditions?

  2. Given that different model families use different input representations, could the authors clarify to what extent the reported performance differences arise from model architectures rather than differences in the available input information?

3 comments

Thank you for your review. We appreciate that you recognized the value of the dataset and benchmarking protocol, as well as the analysis across prediction horizons and delay-change regimes.

W1/Q1(A)

We agree that the sensitivity analysis does not quantify the accuracy of inferred routes. We therefore checked how often a plausible alternative exists and found one for only 0.40% of event pairs requiring route inference in January 2025. A plausible alternative has to connect the same consecutive event locations, match the departure and arrival lines recorded in the raw data, and contain no loops or backtracking. January 2025 contains 2,820,828 events, with 1,856,170 consecutive-event pairs requiring route inference. The search produced 44 distinct candidate cases. Manual inspection confirmed credible alternatives in 33 of them, accounting for 7,423 pairs (0.40%). Thus, the route we inferred was almost always the only plausible one consistent with the recorded lines.

W1/Q1(B)

Without ground-truth paths, we cannot measure the MAE effect of errors in our inferred routes, but our ablation shows that features derived from those routes improve prediction accuracy. In the lite-tier MLP ablation (Appendix C.6.7, Table 45), removing this context raises test MAE from 79.73 to 81.20 seconds (+1.47). This is the third-largest increase, after removing past delay features (+4.26) and planned timing (+3.83). The ablation shows that features derived from inferred routes improve prediction, but it does not measure the effect of incorrect routes.

W2/Q2

We agree that a more systematic comparison would clarify RIDE's contribution and provide one below using the same properties for each cited dataset.

PropertyChina (Zhang)Italy (Wu)Germany (Döllmann)Finland (Borin)RIDE (Ours)
Period16 weeks6 months4 years (yearly updated)8 years3 years
Network scopeHigh-speed railway networkNational passenger networkNational passenger networkLong-distance servicesNational passenger network
Processed dataTrain timings, station connections, weatherTrain timings, station connections, weatherTrain timings, station metadataTrain timings, station metadata, weatherTrain timings, network topology, weather
Model-specific datasetsNoNoNoNoYes
Train/test splitNoNoNoNoYes
Evaluation protocolNoNoNoNoYes

RIDE combines national passenger-network coverage, three years of train events, detailed network topology, and weather data, a combination none of the cited datasets provides. This supports a full-year test set across the national passenger network and methods that use infrastructure and weather information. Unlike the other cited datasets, RIDE provides fixed benchmark splits and a unified evaluation protocol, allowing methods to be compared on the same prediction targets. We will add this comparison to the paper.

Q3

The relative model rankings are mostly stable on the test set, with the GNN best in every month and most train categories, apart from a few less frequent categories. Across months, the model ranking is unchanged except that the LSTM and Transformer switch places in some months, which is expected given their close aggregate MAE. Across the ten train categories, the GNN ranks first or a close second in eight. Surprisingly, XGBoost leads in ICE and CHARTER, perhaps because these categories are much smaller, although we have not established the cause. Thus, the aggregate results reflect broadly stable rankings across months and most train categories, although absolute errors vary by month and some smaller categories favor other models. We will add these breakdowns to the paper.

Q4

We acknowledge that the models use different input representations. Each representation is chosen to match how its architecture organizes and combines information, and is therefore part of the method being evaluated. But all learning-based models use the same data sources and feature families. What differs is how this information is fed into each model: the MLP and XGBoost have one vector per train, the LSTM arranges this vector into static features and sequence features, the Transformer inputs each train's vector representation as a token and can propagate information through attention, while the GNN distributes these features on a graph and uses message passing to propagate information. For example, the tabular models put the distance, train count, and average delay of each of the next 10 links into the train’s feature vector. The GNN puts those quantities directly on the corresponding links in its graph. The Transformer and GNN can also use information from other trains beyond the local summaries available to the per-train models. These specificities are part of how each architecture works, so removing them would change the methods being compared.

Thank you for the authors' response. We believe their responses have largely addressed our concerns regarding Q1, Q3, and Q4. Regarding Q2, we still have the following reservations.

We acknowledge the value of RIDE's unified evaluation protocol in facilitating fair comparisons. However, researchers could also establish standardized evaluation protocols across multiple existing datasets, enabling fair comparisons within each dataset while further assessing models' cross-dataset generalization. Such multi-dataset evaluation is also valuable.

Therefore, we consider RIDE's primary contribution to be complementing existing data resources: it integrates multiple data attributes into a single dataset, whereas multiple existing datasets would need to be combined to cover these dimensions. However, the authors have not sufficiently clarified what additional value this data integration and unified evaluation design offer compared with an evaluation framework built upon multiple existing datasets. For these reasons, we maintain our original score.

Thank you for your response.

We agree with your additional comment that extending standardized evaluation across multiple railway datasets would be valuable, particularly for studying cross-network generalization. However, developing such a benchmark would require substantial additional work. Existing datasets lack comparable network topology and would require country-specific processing, harmonization of event records, and validation of data quality and coverage. These steps would require considerable effort for each additional network, making this a direction for future work.

While a multi-dataset benchmark would offer additional evaluation capabilities, we believe RIDE should be viewed as a contribution towards this broader objective, rather than as a competing alternative. RIDE establishes a foundation for this future direction by providing a standardized representation of railway operations, a snapshot-based framework defining common prediction targets, and a unified evaluation protocol, together with the processing pipeline and baseline implementations needed to support them. Extending this framework to other networks would allow researchers to evaluate cross-network generalization while retaining the reproducibility and comparability that RIDE currently provides within the Belgian network. We therefore view a multi-dataset benchmark as a natural continuation of RIDE rather than an alternative to its contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.