Mohd Zamin Quadri

GitHubLinkedIn

All work

Master's thesis / Reliable ML

Reliable GNN Surrogates for Transport Policy Analysis

An evidence-led study of when a graph neural network surrogate is accurate, when it is uncertain, and how that uncertainty can support review decisions.

Role
Researcher and thesis author
Classification
Academic research
Institution
Technical University of Munich
Inspect repository
Information spreads through a fixed road network, then each junction carries how unsure the surrogate is there.A road network of twenty-five junctions. One junction is intervened on and the effect propagates outward along real hop distance; the network then lifts so that height carries that distance, a prediction appears at each junction, and a band of uncertainty grows around it with distance from the intervention. A reliability inset shows the tracked calibration curve moving from 0.2687 to 0.0479 expected calibration error, and the least confident junctions are finally marked for human review.

01Problem

Why this work exists

Paris-scale transport simulations are expensive, while a fast surrogate can fail unevenly across road links and policy regimes. Point predictions alone do not tell an analyst when to trust the approximation.

02Contribution

What I can claim

Built on the MATSim corpus and PointNetTransfGAT infrastructure of prior work, then evaluated training variants, MC Dropout, deep ensembles, sigma scaling, split and adaptive conformal prediction, selective prediction, CQR, and error-detection diagnostics.

Technical questionHow does an expensive transport simulation become a confidence-aware review decision?
  1. 01SimulateCapacity-reduction policy scenarios
  2. 02RepresentOne graph node per road link
  3. 03ApproximatePointNet + TransformerConv + GAT
  4. 04QuantifyDropout, ensembles, and intervals
  5. 05DecideCalibrate, retain, or route to review
ArchitectureFive features per road segment, twice through a PointNet against each end of the segment, then attention over the neighbourhood, and a graph convolution for the single number out.
Layer sizes read from scripts/gnn/models/point_net_transf_gat.py, not from the write-up
StageShape
PointNet, start positionLocal MLP 7 to 256, global MLP 256 to 512 to 512
PointNet, end positionLocal MLP 514 to 256, global MLP 256 to 512 to 128
TransformerConv128 to 256, four heads
TransformerConv256 to 512, four heads
GATConv512 to 64
GATConv64 to 1
Technical questionOn the same five features and the same held-out scenarios, does the graph surrogate predict better than an ordinary tabular model?

No. A gradient-boosted tree reaches R² 0.7414 against the best graph model’s 0.6841, and trains in about three minutes rather than hours. The graph model earns its place a different way: on ranking its own errors, MC dropout on the weaker network reaches Spearman 0.4817 against the ensemble’s 0.3997, which is what the selective-review result is built on. The tree carries no notion of the network at all.

All rows: 80/10/10 scenario-level split, seed 42, 100 held-out scenarios, 3,163,500 test nodes. Source: results/trials/non_gnn_baseline_results.json
ModelMAE (veh/h)RMSE (veh/h)
XGBoost0.74142.77395.6933
Deep ensemble0.68413.48536.2927
Random forest0.66123.26286.5164
T8 with MC dropout0.58563.94797.2073
MLP0.49283.88317.9730
What did not workEight runs, in order. The two that weight the loss toward the large changes are the clearest result in the table, and neither was pursued.
Source: docs/verified/VERIFIED_RESULTS_MASTER.csv
RunWhat changedSplit
T1First runExcluded from comparison: a linear final layer, not the graph convolution T2 onward use.not comparable80/15/5
T2Correct architecture0.511780/15/5
T3Weighted lossWeighting the loss toward large changes halves the fit.0.224680/15/5
T4Weighted loss, dropout back0.242680/15/5
T5Smaller batch0.555380/15/5
T6Lower learning rate0.522380/15/5
T780/10/10 split0.547180/10/10
T8Lower dropoutThe model every uncertainty result on this page is built on.0.595780/10/10
Technical questionHow does accepted-set error change at six audited review-capacity choices?
Six selected audited operating points; values are observed, not interpolated
RetainedAccepted-set MAEReview queue
10%1.05 veh/h2,847,150
25%1.79 veh/h2,372,625
50%2.32 veh/h1,581,750
75%2.80 veh/h790,875
90%3.23 veh/h316,350
100%3.95 veh/h0

Six selected audited operating points from a 5-percentage-point evaluation grid. The interface never interpolates between them and shows no random-review baseline. Rows outside the retained set enter a review queue; the analysis does not label them wrong or unsafe.

Method

How a fast approximation earns the right to be used

An agent-based transport simulation is expensive enough that policy questions get asked less often than they should. A surrogate answers in a fraction of the time - and a surrogate that is confidently wrong is worse than the simulation it replaced. This is the sequence that closes that gap.

  1. Represent

    The network is the input. A road link's behaviour depends on what it connects to, so the model reads a graph rather than a table of independent rows.

  2. Propagate

    Information reaches a link by travelling through the structure. The wavefront here is a real breadth-first traversal of the fixed network, so the order is the graph's, not a designer's.

  3. Predict

    Every junction now has an estimate. On its own that is a number with no indication of when to believe it.

  4. Quantify

    Each junction gets a measure of how much the model had to work with. The further information had to travel, the less there was to go on.

  5. Decline

    The junctions the model is least sure about are marked rather than hidden, which is what makes a review queue possible instead of a blanket trust decision.

  6. What this does not show

    The figure is a construction, not a measurement: ring size is a function of hop distance in this drawing. The measured relationship between uncertainty and error is in the evidence table below, along with what it does not establish.

03System

Workflow and decisions

The research separates fast surrogate inference from uncertainty estimation, calibration, and the downstream accept-or-review decision. Each layer has a distinct evaluation protocol so confidence claims are not inferred from point accuracy alone.

  1. 01MATSim policy scenarios
  2. 02Road-link graph representation
  3. 03PointNet + TransformerConv + GAT surrogate
  4. 04Uncertainty estimation
  5. 05Calibration and review policy
  • PyTorch
  • PyTorch Geometric
  • NumPy
  • SciPy
  • scikit-learn
  • Streamlit
  • LaTeX

04Evidence

What is actually versioned

Held-out scope100 scenarios

3,163,500 road-link predictions in the cached test artifacts.

Primary GNNR² 0.596

Deterministic Trial 8; MAE 3.96 veh/h and RMSE 7.12 veh/h.

Uncertainty rankingρ 0.482

Pooled Spearman correlation between MC Dropout uncertainty and absolute error.

Selective review41.2% lower MAE

Accepted-set MAE at 50% retention versus accepting every prediction.

Deep ensembleR² 0.684

5-member ensemble; MAE 3.49 veh/h and uncertainty-error ρ 0.400.

Marginal coverage90.02% / 95.01%

Reported final-thesis split-conformal protocol at nominal 90% / 95%.

05Inspection points

Go directly to the evidence

06Quality controls

How the work is checked

  • Prediction-to-analysis workflow backed by tracked numeric artifacts
  • Aggregate-only local evidence dashboard with automated regression checks
  • Calibration protocols versioned instead of combining incompatible splits
  • Thesis claims cross-checked against generated reports and cached predictions

07Limitations

Where the evidence stops

  • One Paris network, one capacity-reduction intervention family, and a fixed 1,000-scenario subset
  • Raw MATSim data is not publicly redistributable; a fresh clone cannot reproduce raw simulation-to-graph processing
  • Historical preprocessing used split-specific scalers, creating an evaluation-distribution methodology risk
  • Uncertainty ranking weakens in some high-change regimes, and coverage is marginal rather than a per-scenario guarantee
  • The portfolio application analyzes cached artifacts; it is not a live policy simulator
  • On the same five features and the same held-out split, a gradient-boosted tree reaches R-squared 0.7414 against the best graph model's 0.6841. The graph surrogate is retained for its uncertainty behaviour, not for its accuracy.

What this changed in my practice

Reliable ML is not one score. Ranking, calibration, conditional behavior, compute cost, and the operational cost of review must be evaluated together.

Next evidence milestone: A future replication should fit preprocessing only on training data and test transfer across networks and intervention families.