## Composite Visualization: Super GNN Training & Prediction Analysis ### Overview The image contains four subplots arranged in a 2x2 grid, analyzing the training and prediction performance of a Super GNN model. Each subplot provides distinct insights into loss dynamics, validation metrics, precision-recall trade-offs, and score distributions. --- ### Components/Axes #### Top-Left: Training Loss Evolution - **Title**: Training Loss Evolution - **Y-Axis**: Loss (0.7 to 1.4) - **X-Axis**: Epoch (0 to 5000) - **Legend**: Blue line labeled "Training Loss" - **Trend**: Loss starts at ~1.38, drops sharply to ~0.72 by epoch 1000, then gradually declines to ~0.71 by epoch 5000. #### Top-Right: Validation Metrics - **Title**: Validation Metrics - **Y-Axis**: Score (0.7 to 0.95) - **X-Axis**: Epoch (0 to 5000) - **Legend**: - Blue circles: "AUC" (starts at 0.75, peaks at ~0.95 by epoch 1000) - Orange squares: "AP" (starts at 0.77, peaks at ~0.95 by epoch 1000) - **Trend**: Both metrics rise sharply initially, plateauing near 0.95 with minor fluctuations. #### Bottom-Left: Precision at Different Recall Levels - **Title**: Precision at Different Recall Levels - **Y-Axis**: Precision (0 to 1) - **X-Axis**: Recall Level (-0.2 to 1.2) - **Color Gradient**: Blue (darker = higher precision) - **Key Data**: - Highest precision (~1.0) at recall 0.0 - Precision decreases as recall increases (e.g., ~0.9 at recall 0.2, ~0.85 at recall 0.4). #### Bottom-Right: Super GNN Score Distribution - **Title**: Super GNN Score Distribution - **X-Axis**: Prediction Score (0.94 to 0.965) - **Y-Axis**: Count (0 to 16) - **Legend**: Red dashed line at 90th percentile (0.954) - **Distribution**: - Peak count (~14) at ~0.945 - Counts decrease toward edges (e.g., ~2 at 0.965) - 90th percentile marked at 0.954. --- ### Key Observations 1. **Training Loss**: Initial instability (spike to 1.38) followed by rapid convergence to stable loss (~0.71). 2. **Validation Metrics**: Both AUC and AP improve rapidly, stabilizing near 0.95, indicating strong generalization. 3. **Precision-Recall Trade-off**: High precision at low recall, but precision degrades as recall increases. 4. **Score Distribution**: Most predictions cluster tightly around 0.95, with a long tail toward higher scores. --- ### Interpretation - **Training Dynamics**: The sharp initial drop in loss suggests effective early learning, while the plateau indicates model stabilization. - **Validation Performance**: High and stable AUC/AP scores confirm the model’s robustness on validation data. - **Precision-Recall Balance**: The heatmap highlights a trade-off: optimizing for recall reduces precision, critical for applications requiring balanced performance. - **Score Distribution**: The concentration of scores near 0.95 suggests consistent model confidence, with the 90th percentile (0.954) indicating strong upper-bound performance. The analysis reveals a well-trained model with effective generalization, though the precision-recall trade-off warrants further tuning for specific use cases.