Beyond the Graph: An Adaptive Meta-Learner Fuses Explainability, Weather, and Dynamics for Robust Bus ETA Prediction

1 Introduction

• Data sparsity and limited capture of complex urban traffic dynamics remain challenges that can make bus arrival estimates unreliable, affecting passenger satisfaction and ridership.The paper frames accurate arrival prediction as important to reliable public transportation and sustainable urban mobility.Source: §1 · p. 2
• The study proposes a hybrid framework for Kolkata’s bus system that integrates diverse data sources with graph-based convolutional networks and neural operators.Its stated aims include a robust spatiotemporal network model and a dynamic hybrid ensemble.Source: §1 · p. 2
• The authors hypothesize reduced prediction errors, better modeling of nonlinear traffic patterns than traditional baselines, and superior performance across varied operating scenarios.These are stated as hypotheses H1, H2, and H3, respectively.Source: §1 · p. 2

2 Related Work

• Statistical approaches modeled recurring patterns, temporal trends, uncertainty, or noisy data, but faced limits in nonlinear dynamics, spatial interactions, computation, or generalizability.Examples include seasonal spatiotemporal modeling, time-series analysis with gradient boosting, Bayesian prediction, and hypercube clustering.Source: §2 · p. 2
• Machine-learning work addressed baselines, incomplete trajectories, and feature engineering, yet retained gaps in spatial dependencies, path quality, graph modeling, or privacy.The cited examples include boosting, route recovery, feature-engineered and classical models, and ANN pipelines.Source: §2 · p. 3
• Deep-learning methods learned complex temporal patterns, but reported challenges included long-range dependencies, external-data reliance, unmodeled road structure, privacy, route-specific models, and engineering complexity.The passage surveys sequence models, traffic-forecast combinations, and other deep architectures.Source: §2 · p. 3
• Graph-based methods directly modeled transportation networks and long-range dependencies, while scalability, computational efficiency, and inference cost remained concerns for real-time use.The cited approaches use graph convolution and hybrid graph embeddings [28], with attention mechanisms in Reich et al..Source: §2 · p. 3
• Hybrid frameworks combine paradigms or data sources and can improve robustness, but their complexity, training demands, and interpretability can impede practical deployment.The passage discusses HSETA, a Wide-Deep recurrent hybrid, and hybrid ensembles.Source: §2 · pp. 3–4

2.1. Summary of Literature and Proposed Solutions

• The proposed work targets gaps in spatial discretization, nonlinear temporal patterns, weather handling, periodic spatiotemporal behavior, sparse data, real-time assimilation, explainability, and computational demands.It presents its methodologies as addressing these gaps while seeking improved prediction accuracy and deployment feasibility.Source: §2.1 · p. 4

3 Research Methodology

• The methodology has three phases: data acquisition and preprocessing, spatiotemporal network construction, and hybrid prediction.The phases establish inputs, map transport dynamics, and perform inference, respectively.Source: §3 · p. 4
• The design integrates high-frequency, granular bus-operation data with contextual environmental factors for bus arrival time prediction.The paper identifies the complex interplay of vehicle paths, recurrent traffic patterns, and stochastic factors such as meteorological conditions as a central modeling challenge.Source: §3 · p. 4
• The hybrid framework is intended to combine mechanistic, statistical, and deep-learning paradigms for robustness and accuracy across dynamic operating conditions.This is the study’s stated rationale for the integrated design.Source: §3 · p. 4

3.1 Data Acquisition and Preprocessing

• The data sources include geo-referenced bus-stop locations and hourly weather records for temperature, humidity, precipitation, wind speed, and cloud cover.Bus stops serve as network nodes and prediction targets; weather parameters are treated as traffic-flow and arrival-time covariates.Source: §3.1 · p. 5
• GPS preprocessing calculates speed between successive points and discretizes coordinates into approximately 50 m × 50 m grid cells, using median coordinates to define stable nodes.The passage states that the grid reduces computational complexity while retaining geographic fidelity and that median coordinates mitigate GPS noise.Source: §3.1 · p. 6
• GPS trajectories were collected from August 2024 to March 2025 on three Kolkata routes: DN-2/1, KB-22, and KB-16.The routes connect Shapoorji with Dakshineswar, Kolkata Station, and Bangoor, respectively.Source: §3.1 · p. 7
• The study-period collection was broadly even, with slight weighting toward monsoon months August–October and a total distribution of 100%.Table 3 attributes the greater monsoon collection frequency to the focus on weather covariates.Source: Table 3 · p. 8

3.2 Spatiotemporal Network Construction

• A Minimum Spanning Tree is extracted from a distance-weighted complete graph to approximate the road network’s sparse topology.The complete graph connects grid-cell nodes using Haversine distances between their median coordinates.Source: §3.2 · p. 8
• Dijkstra’s algorithm finds the shortest geographic path from a bus location to its target stop on the MST-augmented graph.That path defines the expected route sequence used by subsequent ETA models.Source: §3.2 · p. 8
• The graph starts as a complete graph of grid cells whose edge weights are Haversine distances between cell median coordinates.The grid cells form the vertices of the directed transportation graph.Source: §3.2 · p. 8

3.3.1 Model 1: MST-AV (Historical Average Baseline)

• MST-AV provides a non-temporal historical baseline, estimating path-edge travel times from mean speeds at the edge’s incident nodes.Historical mean speeds are aggregated for each network node and averaged across an edge’s endpoints.Source: §3.3.1 · p. 9

3.3.2 Model 2: GDRN-DFT (Spectral Pattern Analysis)

• GDRN-DFT diffuses and imputes missing network speeds, models temporal dependencies, then uses DFT to identify periodic components and forecast speeds.Its training loss focuses on discrepancies at observed locations; inverse DFT yields smoothed future speeds for ETA calculations.Source: §3.3.2 · p. 10

3.3.3 Model 3: KNN (Temporal Dynamics via Koopman Theory)

• The Koopman Neural Network lifts nonlinear speed sequences into a latent space where a Koopman matrix approximates linear evolution, then reconstructs future predictions.The encoder creates the finite-dimensional representation and the inverse mapping returns predictions to the observable data space.Source: §3.3.3 · pp. 10–12

3.3.4 Model 4: FENN (Feature-Engineered Forecasting)

• FENN encodes sequences of location, weather, periodic-time, and speed features, then uses a feedforward network to predict future transformed speeds.Training minimizes squared error between actual and predicted transformed speeds.Source: §3.3.4 · pp. 12–13

3.3.5 Model 5: MGCN(Real-Time Graph Learning)

• MGCN combines observed speeds with prior estimates through a binary mask, then applies graph convolutions to predict the next-step speed map.Its composite loss balances accuracy on observed data with consistency against prior estimates for unobserved data.Source: §3.3.5 · pp. 13–14

3.3.6 Model 6: HYB(nm)(Hybrid Framework)

• HYB(nm) uses a context-aware meta-learner to assign blending weights to five base models and combine their ETA predictions.The meta-learner maps location, weather, periodic time, current speed, and distance to the target into weights; downsizing can retain only the top nm models.Source: §3.3.6 · pp. 14–15
• HYB(nm) can retain only the top nm model weights, renormalize them, and calculate its ETA as a weighted sum of base-model ETAs.The meta-learner’s loss combines a weight-summation constraint with squared error between the ensemble ETA and actual ETA.Source: §3.3.6 · pp. 15–16

4.1 Evaluation Metrics

• MAE, RMSE, and MAPE assess error magnitude, larger-error penalties, and percentage accuracy, respectively; each route’s test set contains 10,000 trips.Assessments use GPS-derived ground-truth arrival times.Source: §4.1 · pp. 16–17

4.2 Performance Comparison Between the Models

• HYB-COMB has the lowest MAE, MAPE, and RMSE across normal, bus-stop-group, and extreme-weather scenarios, while specialists lead in some ranges.KOOP-NET excels at short-range normal-condition predictions, FE-NN at short-range extreme-weather predictions, RT-GCN at medium distances, and GDRN-DFT at longer horizons; errors generally increase in extreme weather.Source: §4.2 · pp. 17–19

4.3 Downsizing Effect of Hybrid Model

• HYB(2) records MAPE of 12.3–14.9%, MAE of 4.5–5.7 minutes, and RMSE of 5.4–6.9 minutes across evaluated routes.It is reported as consistently lowest or near-lowest in MAPE, particularly strong for 0–10-minute ETAs, and competitive at longer horizons; HYB(1) occasionally has the lowest absolute errors.Source: §4.3 · pp. 19–20, Table 5 · p. 19

4.4 Comparison with Baseline Models

• DCRNN, MTGNN, and HYB(2) lead short-range forecasting, while GWNet, HYB(2), and MTGNN are competitive at longer horizons.ST-ResNet and ST-GConv are less effective on irregular graphs; model selection should account for ETA range and graph topology.Source: §4.4 · pp. 21–24
• 48.9% is HYB(2)’s highest overall win rate across baselines, routes, ETA ranges, and MAE, MAPE, and RMSE.Table 8 defines win rate as the percentage of ETA ranges where a model records the lowest error; the reported overall percentage is route-averaged.Source: §4.4 · p. 24, Table 8 · p. 24

4.5 Hyper Parameter Study

• All models are sensitive to hyperparameters, especially temporal resolution and architectural depth, with tuned configurations balancing accuracy and computational efficiency.The study uses grid search over discrete ranges and reports that optimal settings avoid underfitting and computational explosion.Source: §4.5 · pp. 25–26

5 Conclusion

• HYB(2) is reported to deliver top accuracy, stability, and efficiency across the tested conditions and routes.The conclusion says its performance matches or surpasses leading models less suited to unpredictable bus networks.Source: §5 · p. 27
• The conclusion assigns distinct roles to individual models: MST-AV is fast and interpretable, KOOP-NET suits short-term predictions, FENN handles sudden weather changes, and MGCN combines real-time traffic with network structure.FENN is described as requiring significant data; GDRN-DFT is presented as balancing spatial and temporal accuracy.Source: §5 · p. 27
• HYB(k) lets agencies blend base models to fit their data, computing power, and operational needs, while simpler models remain options for limited-resource systems.The conclusion presents this as a toolkit spanning economical single models and adaptable hybrids.Source: §5 · p. 27

• Evaluation on only three Kolkata routes may under-represent infrastructure and urban morphologies elsewhere, making broader metropolitan validation essential.Source: §6 · p. 27
• Using GPS and weather data omits incidents, obstructions, and passenger loads that may critically affect predictions.Source: §6 · p. 27
• The authors call for refining preprocessing, loss weighting, regularisation, and latency, alongside context-adaptive model selection and multimodal data integration.They also propose efficient ingestion, model benchmarking and tuning, and edge-deployable meta-learners with caching.Source: §6 · p. 27

• GPS points are mapped to a 50m × 50m grid and used to build a graph, minimum spanning tree, adjacency matrix, and shortest paths; weather data are also acquired.Source: Appendix A · p. 29
• MST-AV computes historical mean speeds for graph nodes and uses them along a shortest path to estimate arrival time.Source: Appendix A · pp. 29–30
• GDRN-DFT imputes masked speed observations with a bidirectional LSTM and derives DFT parameters after filtering low-amplitude patterns.Source: Appendix A · p. 30
• The Koopman neural network trains an encoder, Koopman matrix, and decoder on speed sequences using reconstruction and prediction losses.Source: Appendix A · p. 32
• FENN encodes location, weather, cyclical time features, and speeds, then predicts future speeds for iterative ETA computation.Source: Appendix A · pp. 34–35
• MGCN training uses observed and DFT speeds with a mask, while real-time ETA computation uses predicted speeds when observations are masked.Source: Appendix A · pp. 36–37
• HYB(nm) trains a neural network to produce context-dependent model weights, then retains and renormalizes the top nm weights to combine component ETAs.Source: Appendix A · p. 38
• The appendix also specifies hybrid ETA computation and forward-pass procedures for benchmark spatiotemporal graph models.Source: Appendix A · pp. 38–39

• Tables B1 and B2 present Route 2 and Route 3 performance analyses using MAE, MAPE, and RMSE.Source: Table B1 · p. 40, Table B2 · p. 40
• Table B3 studies MGCN MAE and computation time across history length, sampling interval, and layer count.Source: Table B3 · p. 41
• Table B4 reports MST-AV performance across different grid sizes.Source: Table B4 · p. 41