David's architecture for his scripts - davidlabee/Graph4Air GitHub Wiki

Aspect	graph4air_workflow.py	graph4air_hyperparammodelsearch.py	graph4air_hyperparamgraphaugmentationsearch.py
Metrics packages	`mean_squared_error`, `mean_absolute_error`, `r2_score`	`mean_squared_error`, `r2_score`, `mean_absolute_error`	`mean_squared_error`, `mean_absolute_error`, `r2_score`
Internal validation	Train op `data.train_mask`, valideren op `data.test_mask`	Optuna-trials trainen op `data.train_mask`, valideren op `data.test_mask`	Train op `data.train_mask`, valideren op `data.test_mask`
validation	Evaluatie op `data.test_mask`	Evaluatie (RMSE) op `data.test_mask` binnen elke trial	Evaluatie (`eval_rmse`) op `data.test_mask`
Train/test split	`valid_idx = np.where(~np.isnan(y))[0]`; shuffle; 80/20; boolean-masks	`valid_idx = np.where(~np.isnan(y.flatten()))[0]`; shuffle; 80/20; boolean-masks	`valid_idx = np.where(~np.isnan(y.flatten()))[0]`; shuffle; 80/20; boolean-masks
Hyperparam tuning	–	Optuna over: `num_layers` (2–4), `hidden_units` {16,32,64}, `dropout` [0–0.5], `lr` [1e-4–1e-2], `weight_decay` [1e-5–1e-3], `activation` {relu,elu}	Optuna over: `top_n` [500–1500], `neighbors` [10–200], `sim_thresh` [0.80–0.9999], `min_dist` [50–500], `max_dist` [500–5000], `hop_thresh` [1–5], `max_edges` [500–5000], `per_node_cap` [1–10]
Epochs & Early stopping	GCN & GAT: 200 epochs, geen early stopping	Elke trial: 100 epochs,wel early stopping	Fase 1: 50 epochs; fase 2: 200 epochs; wel early stopping
Node masking	`valid_idx = indices waar y≠NaN`; excludes outliers; boolean-masks	`valid_idx = indices waar y≠NaN`; boolean-masks	`valid_idx = indices waar y≠NaN`; boolean-masks