🗓️ Meeting Notes – 29 April 2025, 14:00

👥 Attendees

• David
• Pieter
• Zhendong
• Jules

---

🧠 Discussion Points

📌 David’s Presentation – Subject Overview and Edge Testing

• Summary:
  • David first introduced his thesis topic to Jules, explaining the goal of enhancing road network graphs for better air pollution modeling.
  • He presented five different augmentation strategies developed and tested (road_graph_strategies_subset-2.ipynb).
• Key Insights:
  1. Feature category-based selectivity (cosine similarity on traffic, population, land use, morphology).
  2. Sparse multi-modal agreement (≥3 feature domains agreeing).
  3. Distance-constrained similarity edges (500 m–20 km range).
  4. Top-K strongest similarity edges (K=2).
  5. Soft augmentation with similarity as edge weights.
  • Strategy 2 (Sparse Multi-Modal Agreement) currently shows the best balance between sparsity and meaningful connectivity.
• Challenges:
  • Avoid over-densifying the graph.
  • Maintain functional similarity while preserving spatial structure.

📌 Pieter’s Presentation – Baseline GAT and GCN Models

• Summary:
  • Pieter walked through the baseline models (Baseline_50m_Thesis_(COPY).ipynb).
  • GAT and GCN were trained on the raw 50 m-segment road graph without augmentation.
• Key Insights:
  • GAT already outperforms GCN in initial experiments.
  • Likely reason: GAT uses attention mechanisms to weigh more important neighbors more heavily, improving information aggregation.
• Challenges:
  • Further analyze why attention benefits this context.

📌 Discussion – Train/Test Splitting Strategy

• Summary:
  • Debated how to best split data for training and evaluation.
• Key Insights:
  • Traditional random splits may or may not be ideal.
  • Our goal is interpolation across known road segments, not predicting truly unseen segments.
• Challenges:
  • Decide if a held-out test set is necessary.
  • Alternatives: cross-validation over segments, internal validation (early stopping), etc.
  • Emphasize achieving smooth, accurate visual maps over raw accuracy.

📌 Additional Topic – Graph Aggregation Alternatives

• Summary:
  • Jules and Pieter met earlier to discuss alternative node aggregation methods.
• Key Ideas:

  • Square-grid aggregation: group segments into grid cells as nodes. This could function as a layer of nodes that captures patterns only observed at a course level. Later this grid layer and the 50m segment layer could be combined in an hierarchical GNN model that captures both low and high resolution patterns.

  • Multi-graph approach: build separate graphs by functional categories (e.g., residential vs. highways). Suggested by Zhendong as faster to implement.

• Challenges:
  • Time overhead of square-grid aggregation.
  • Deciding how to train/evaluate multiple specialized graphs.

---

📅 Next Steps

• David to benchmark and finalize the best augmentation strategy and test it on the whole city (Sparse Multi-Modal Agreement).
• Pieter to merge the current 50m segments into 100m and 200m segments (possibly adding segment length as a feature if segment length varies a lot)
• Group to:
  • Describe baseline model in the wiki while adressing the challenges and motivating choices.
  • Define final evaluation methodology (Basic holdout vs. cross-validation)(with or without early stopping?).
  • Look into using multiple graphs. So a graph for residential areas, highways etc.

Next Meeting:
6 may 2025, 14:30