Scoreboard - davidlabee/Graph4Air GitHub Wiki

GNN Experiment Scoreboard

This table helps track and compare results between team members for each graph construction and model configuration. Please fill in the results as you test new combinations. Metrics are displayed for internal and external validation.

Comparison to Data-Only (internal validation)

Contributor Graph Model Layers Epochs MAE RMSE Pearson R² Hyperparameters (hu's, lr, dropout, activ.(, heads)) #runs
Baseline original 50m GCN 4 early stopping 4.44 6.54 0.710 128, 0.020, 0.4, elu 10
Baseline original 50m GAT 5 early stopping 4.09 6.24 0.739 128, 0.012, 0.0, elu, 1 10
-- -- -- -- -- -- -- -- -- --
Pieter hierarchical-4 GCN 3 200 5.75 8.10 0.55 32, 0.010, 0.0, relu 5
Pieter hierarchical-4 GAT 3 200 5.54 7.81 0.58 32, 0.010, 0.0, relu, 1 5
Pieter hierarchical-8 GCN 3 200 5.72 8.07 0.56 32, 0.010, 0.0, relu 5
Pieter hierarchical-8 GAT 3 200 5.52 7.82 0.58 32, 0.010, 0.0, relu, 1 5
Pieter hierarchical-16 GCN 3 200 5.75 8.12 0.55 32, 0.010, 0.0, relu 5
Pieter hierarchical-16 GAT 3 200 5.63 7.96 0.57 32, 0.010, 0.0, relu, 1 5
Pieter hierarchical-32 GCN 3 200 5.86 8.25 0.54 32, 0.010, 0.0, relu 5
Pieter hierarchical-32 GAT 3 200 6.08 8.55 0.50 32, 0.010, 0.0, relu, 1 5
-- -- -- -- -- -- -- -- -- --
David graph augmentation optimised hyperparameters GCN 4 200 5.59 8.04 0.58 64, 1
David graph augmentation optimised hyperparameters GAT 4 200 4.99 7.38 0.63 64 1
David graph augmentation optimised hyperparameters GCN 4 1000 (500) 5.248 7.632 0.619 5-fold 5
David graph augmentation optimised hyperparameters GAT 4 1000 (350) 4.908 7.2 0.661 5-fold 5

Comparison to Palmes Tubes (external validation)

Contributor Graph Model Layers Epochs MAE RMSE Hyperparameters #runs
Baseline original 50m GCN 3 200 4.21 5.30 ... 64, 0.010, 0.5 #runs
Baseline original 50m GAT 3 200 3.71 4.81 ... 64, 0.010, 0.2 #runs
-- -- -- -- -- -- -- -- -- --
Pieter hierarchical-4 GCN 3 200 4.63 5.94 0.53 32, 0.010, 0.0 5
Pieter hierarchical-4 GAT 3 200 4.91 6.37 0.51 32, 0.010, 0.0 5
Pieter hierarchical-8 GCN 3 200 4.45 5.63 0.55 32, 0.010, 0.0 5
Pieter hierarchical-8 GAT 3 200 4.78 6.09 0.53 32, 0.010, 0.0 5
Pieter hierarchical-16 GCN 3 200 4.53 5.69 0.53 32, 0.010, 0.0 5
Pieter hierarchical-16 GAT 3 200 4.79 6.19 0.47 32, 0.010, 0.0 5
Pieter hierarchical-32 GCN 3 200 4.50 5.62 0.51 32, 0.010, 0.0 5
Pieter hierarchical-32 GAT 3 200 5.02 6.35 0.48 32, 0.010, 0.0 5
-- -- -- -- -- -- -- -- -- --
Pieter hierarchical-4 + tuning GCN 4 200 4.91 6.55 0.51 128, 0.010, 0.2, 4 5
Pieter hierarchical-4 + tuning GAT 4 200 5.02 6.72 0.48 128, 0.010, 0.0, 4 5
Pieter hierarchical-8 + tuning GCN 4 200 4.31 5.58 0.54 128, 0.007, 0.35, 4 5
Pieter hierarchical-8 + tuning GAT 5 200 4.27 5.55 0.49 64, 0.016, 0.1, 5 5
-- -- -- -- -- -- -- -- -- --
David graph augmentation optimised hyperparameters GCN 4 200 4.06 5.25 0.60 64, 1
David graph augmentation optimised hyperparameters GAT 4 200 4.24 5.61 0.55 64 1
David graph augmentation optimised hyperparameters GCN 4 1000(500) 3.95 5.28 0.56 64, 5
David graph augmentation optimised hyperparameters GAT 4 1000(350) 4.41 5.98 0.52 64 5