Meeting notes week 10 - davidlabee/Graph4Air GitHub Wiki

🗓️ Meeting Notes – 18 June 2025, 10:15

Pieter created a final workflow which we both ran for all our different graphs
- By doing this we made sure that we both followed exactly the same steps with the same reasoning behind it

Hereby the Model optimization and evaluation protocol:

Hyperparameter tuning on 100% of the labeled data:
- 100 trials (40 “startup” + 60 “normal”)
- 3-fold cross-validation per trial (fixed folds)
- Pruning allowed after each fold
- Train each fold up to 20 epochs
With best hyperparameters, 10-fold cross validation, also on 100% of the labeled data:
- Report mean ± std (internal validation)
- Early stopping after 50 epochs without improvement with a maximum of 2500 epochs
Final model training:
- Retrain the model once more on 100% of the labeled data.
- Using the best hyperparameters found in tuning.
- Early stopping after 50 epochs without improvement with a maximum of 2500 epochs
- Predicts once for all labeled and unlabeled segments.
- Predictions are used for external validation and map visualization

Adjusted the density plots again with the LUR model and mixed effect model predictions as well
Davids final scores:

Finished the first 3 chapters of my thesis and created a design for how I want to structure the final chapters: Using the following link the thesis can be reviewed (could also send it as an pdf) (Still have to work on the page structure and not all sources are loaded in correctly yet): https://www.overleaf.com/read/hmyssrdkkcjm#eca374