Understanding Data Influence in Transportation Networks: A Physics-Constrained Influence-Function Framework
Authors: Hamidreza Attari, Yueyue Fan, Johannes Royset
Organizations: Department of Civil and Environmental Engineering, University of California, Davis, CA, USA · Daniel J. Epstein Dept. of Industrial and Systems Engineering, University of Southern California, CA, USA
Understanding how individual training observations influence model predictions is important for developing reliable and interpretable data-driven transportation systems. Existing influence-function methods, however, are primarily formulated for unconstrained learning and do not explicitly account for physical constraints that couple different components of a transportation network. Under such constraints, perturbing a training observation can affect model parameters and predictions only through physically feasible directions, potentially changing how its influence propagates across the network. This paper develops a hard physics-informed influence-function framework for constrained learning, in which training-data influence is characterized within the physically feasible space defined by network constraints. The analytical results reveal how physical constraints alter the magnitude, direction, and spatial distribution of training-sample influence relative to conventional influence measures. Using traffic prediction under flow-conservation constraints as a case study, we show that influential observations identified by the proposed framework differ systematically from those identified by classical influence functions. Through orthogonal decomposition, we quantitatively characterize how much of a perturbation in the training data is propagated through the network via the flow-conservation constraints and how much is removed by the constraint projection. The proposed influence framework provides a physically consistent approach for evaluating training-data influence in constrained transportation learning problems, with potential applications to data diagnostics, sensor evaluation, and targeted data collection.
Figures & tables
Physics treatment
Sensitivity / data perturbation
No sensitivity consideration
Training behavior
Post-hoc attribution
No constraint
Liu et al. (2026) Ma et al. (2015) Fu et al. (2016) Li et al. (2018) Yu et al. (2018) Guo et al. (2019) Wang and Wang (2022)
—
Koh and Liang (2017)
Soft constraint
Márquez-Neila et al. (2017) Dener et al. (2020)
Wang et al. (2022) De Ryck et al. (2024) Daw et al. (2023) Wang et al. (2023b) Lu et al. (2021) Jagtap et al. (2020)
Krasowski et al. (2026)
Hard constraint
Frerix et al. (2020) Huang et al. (2021) Wang et al. (2023a) Tordesillas et al. (2023) Konstantinov and Utkin (2023) Chen et al. (2024) Min and Azizan (2025)
—
This work
Table 1: Related work organized by how physics is handled and how sensitivity is analyzed.
Figure 1: Schematic illustration of the HardNet-Aff framework, adapted from Min and Azizan Min and Azizan (2025) . A standard neural network first maps the input x to an unconstrained prediction fθ(x) . The prediction is then passed through a differentiable constraint-enforcement layer, where the affine constraint set C(x) is defined for the given input and the projection operator P maps the raw output to a feasible point. The final projected output satisfies P(fθ)(x)∈C(x) by construction, while gradients can still be backpropagated through the projection layer during end-to-end training.
Figure 2: HardNetAff-STGNN architecture for traffic-flow prediction. The STGNN backbone learns spatial and temporal traffic patterns from loop detector data, while the hard equality projection layer enforces traffic-flow conservation to produce physics-feasible station-level predictions (figure generated in ChatGPT).
Table 2: Scaled LiSSA hyperparameters used in experiments.
Data alignment
Physical constraint alignment
Model
MAE
RMSE
R2
∣Ay−b∣
max∣Ay−b∣
Unconstrained
267.209
443.807
0.9147
121.95
536.143
HardNetAff-STGNN
263.659
439.798
0.9162
9.9×10−5
5.66×10−4
Table 3: Test set results on 318-link Sacramento PeMS network (540 test windows). Flow units: veh/hr.
Training set
Mean PIP
0.923
Median PIP
0.931
AM peak (07:00-09:00)
0.948
PM peak (16:00-18:00)
0.938
Night (22:00-05:00)
0.891
Table 4: PIP statistics for the HardNetAff-STGNN.
Figure 3: Spatial distribution of Station Sensitivity Scores (SSS) across the Sacramento PeMS network. Yellow indicates high sensitivity; dark blue indicates near-zero sensitivity. The star-shaped pattern centers on the downtown interchange where I-80, I-5, US-50, and CA-99 meet.
Figure 4: Top 30 stations by importance score under three frameworks: classic influence function on the unconstrained model (left), classic influence function on the physics-informed model (middle), and physics-informed influence function on the physics-informed model (right). Each box is a station at its rank, and each line follows the same station from one framework to the next. Colors show the freeway route. Thick flat lines mark stations with the same rank in both frameworks. Dashed lines mark stations that enter or leave the top 30; their rank outside the top 30 is shown at the bottom.
Figure 5: Top 30 stations by (a) the physics-informed influence score (SSS), (b) mean hourly traffic volume, (c) degree, and (d) betweenness centrality. Red markers show the top 30 stations of each measure; blue rings mark stations that are also in the SSS top 30. The table beside each map lists the ranking and the SSS rank of each station.
Unconstrained
Physics-informed
Holdout value
Δ (%)
Holdout value
Δ (%)
Metric
Top-10
Bottom-10
Top-10
Bottom-10
Top-10
Bottom-10
Top-10
Bottom-10
MAE (veh/hr)
383.99
263.84
+52.6
+4
389.83
265.23
+44.2
−1.87
RMSE (veh/hr)
945.92
439.26
+119
+2.1
947.53
438.90
+114.19
−0.76
R2
0.613
0.917
−33.2
≈0
0.6113
0.9166
−32
≈0
Table 5: Impact of holding out the top-10 and bottom-10 SSS stations on the unconstrained and physics-informed models. All metrics are evaluated across all stations. Δ (%) denotes the percentage change relative to the corresponding full model.
Constrained learning has been increasingly applied to various domains to ensure explicit feasibility requirements due to fairness, safety, robustness, regularization, and physics or logic constraints. Understanding how training samples influence the solution (e.g., learned parameters) of constrained learning is crucial for interpretability and robustness. The classical influence function may become unreliable in constrained settings: data perturbations can reshape both the objective and the feasible region, leading to estimates that violate feasibility. In response, we propose the Directional Influence Function (DIF), a new estimator that explicitly incorporates the constraints into influence estimation. DIF formulates the optimality conditions of constrained learning as a variational inequality (VI) and analyzes how perturbing training data affects this VI. We validate DIF in constrained linear regression and demonstrate that it recovers leave-one-out retraining results, whereas IF and penalty-based IF exhibit significant bias. We further apply DIF to fairness-constrained CNNs, where DIF accurately predicts test loss changes under data removal and aligns closely with actual retraining. Our results establish DIF as an efficient and reliable tool for data attribution in constrained learning.
Xin Wang, R. Tyrrell Rockafellar, Xuegang +1
Jeff · Department of Civil and Environmental Engineering, University of Washington, Seattle, WA, USA
Macroscopic traffic flow is stochastic, but the physics-informed deep learning methods currently used in transportation literature embed deterministic PDEs and produce point-valued outputs; the stochasticity of the governing dynamics plays no role in the learned representation. This work develops a framework in which the physics constraint itself is distributional and directly derived from stochastic traffic-flow dynamics. Starting from an Ito-type Lighthill-Whitham-Richards model with Brownian forcing, we derive a one-point forward equation for the marginal traffic density at each spatial location. The spatial coupling induced by the conservation law appears as an explicit conditional drift term, which makes the closure requirement transparent. Based on this formulation, we derive an equivalent deterministic Probability Flow ODE that is pointwise evaluable and differentiable once a closure is specified. Incorporating this as a physics constraint, we then propose a score network with an advection-closure module, trainable by denoising score matching together with a Fokker-Planck residual loss. The resulting model targets a data-conditioned density distribution, from which point estimates, credible intervals, and congestion-risk measures can be computed. The framework provides a basis for distributional traffic-state estimation and for stochastic fundamental-diagram analysis in a physics-informed generative setting.
Traffic state estimation (TSE) reconstructs traffic density at unmonitored locations from a sparse set of sensors. Data-driven models work well when sensors are dense but degrade sharply as coverage thins. We show that imposing the Lighthill-Whitham-Richards conservation law as a physics-informed prior makes TSE more data-efficient. On four real corridors (NGSIM, I-80, PeMS04, PeMS08), physics-informed estimators generalise to held-out sensors where data-driven models, including classical and deep kriging, overfit the observed ones. On the two shock-rich corridors the effect is a clear crossover: as more sensors are held out, data-driven models collapse while the physics-informed estimator degrades gracefully, leading the no-physics network by about 26-38% at 85-90% held out and the strongest classical kriging by about 10-21% at 90%. A matched smoothness prior does not reproduce the gain, and an extended Kalman filter on the same LWR model is outperformed on the shock corridors, so the advantage comes from how the conservation law is imposed. Finally, one loss-conditional PINN removes the per-corridor loss-weight search, matching a 20-point search at about 5x lower cost.
Anna Lazareva, Alexander Tarakanov
HSE University, Moscow, Russia · Moscow Institute of Physics and Technology (MIPT), Russia · AI VK, Moscow, Russia