Submission of predictions

[emmakowu3579-ui]

Submission

Team Information

• Team Name: emmanuel_owusu
• Model Type: human+llm

Model Description

This submission utilizes a Graph Attention Network v2 (GATv2) architecture built with PyTorch Geometric. The network processes the multi-material graph nodes using two layers of multi-head attention (4 heads) with ELU activations to capture edge-specific importance. A global mean pooling layer aggregates the graph representations, which are then passed through a fully connected regression head.

To stabilize gradients across the vastly different target scales (Pressure, Temperature, and Speed), the pipeline implements strict Z-score Target Normalization. Training was conducted using an 80/20 Train/Validation split, utilizing L1 Loss for backpropagation while actively tracking the Normalized Mean Absolute Error (NMAE) with an Early Stopping monitor (patience=30) to prevent overfitting and capture the best epoch weights.

Checklist

• Added predictions.csv with all required columns (id, pressure, temperature, speed)
• Predictions match test set IDs exactly (120 samples)
• No missing or invalid values
• This is my first and only submission

Notes

The target outputs were strictly inverse-transformed from the normalized 0 to 1 training scale back into their original physical units before generating the predictions.csv file. The id column mapping was directly extracted from the PyG data.batch to ensure perfect ID alignment with the test_nodes.csv.

[github-actions]

Submission Scoring System

[Evaluation Results]

[Combined Score]

NMAE: 0.064701 (6.47%)

Lower is better. This is the official ranking metric.

✅ Score recorded! Check the Leaderboard.

This PR has been automatically closed. Your submission files are NOT stored in the repository.