🗓️ Meeting Notes – 22 April 2025, 9:00

👥 Attendees

• David
• Pieter
• Zhendong

---

🧠 Discussion Points

📌 Pieter’s Presentation – Learning the Right Resolution: A Multi-Scale GNN Approach

• Context: Mobile car-based sensors yield high-resolution NO₂ data at 1s intervals.
• Research Focus: Evaluates how spatial aggregation (10–500m segments) impacts GNN performance.
• Key Questions:
  • What is the optimal spatial resolution for graph construction?
  • Can multi-resolution learning outperform single-scale GNNs?
• Methodology:
  • Construct graphs at multiple scales (25m, 50m, 100m, 200m, etc.)
  • Compare GCN and GAT across resolutions.
  • Propose hierarchical or parallel multi-resolution models.
• Expected Contribution:
  • First mobile-sensor study using multi-scale GNNs for air quality.
  • Performance benchmark across resolutions.
  • A framework for combining multiple resolutions in GNNs.

📌 David’s Presentation – Augmenting Road Segment Graphs with Feature Similarity

• Motivation: Physical adjacency in graphs misses global functional patterns relevant to pollution spread.
• Research Question: Can feature similarity-based edges (e.g. land use, traffic, morphology) enhance GNN predictions while maintaining spatial integrity?
• Graph Design:
  • Base graph: Road network topology (adjacency).
  • Augmentation: Add edges based on cosine similarity (e.g. threshold > 0.9) using feature vectors.
• Cited Studies:
  • Yan et al. (2021), Zhou et al. (2023): Importance of global semantic similarity.
  • Briggs et al. (2020): Global correlations matter for accurate environmental modeling.
  • Madrid Study, UrbanAir, Chen et al. (2022): Showed hybrid/augmented graphs outperform standard spatial-only models.
• Design Challenges:
  • Balancing between preserving spatial locality and introducing functionally meaningful global edges.
  • Risk of graph densification and over-smoothing.

📌 Overall discussion points (feedback from Zhendong)

• Discuss why not use intersections as nodes and back this up with scientific literature.
• Give the presentation more as challenges which there are and how we can tackle them with this research
• For Pieter: A Multi-resolution GNN should have the output on the 50m segments for comparison with other methods
• For David: think theoretically about your suggestion and also look which similarity measure you are gonna use.
• Look at a good baseline model. Zhendong has 2 models (linear model and random forest) to compare the GCN/GAT models. Also find a way to compare with the landuse Regression model. Also look at a baseline graph model that you create yourself.

---

📅 Next Steps

• Both Look at the intersections as nodes and how they could work
• Both Update presentations to include feedback
• Pieter Start experimenting with building the graph
• Pieter Write a mail to Jules to request aggregation of features
• David Start experimenting with building the graph
• David Experiment with different similarity metrics and thresholds for graph augmentation
• Zhendong share the related papers which you mentioned in the meeting
• Zhendong Find a moment for the next meeting

Next Meeting:
NO DATE YET, will have to be determined