## Line Graph: AUC Performance vs. Number of Training Bags ### Overview The graph compares the performance of five machine learning models (measured by AUC) as a function of the number of training bags used. AUC values range from 0.55 to 1.00, with the x-axis representing training bags from 0 to 500. Five distinct lines represent different models, each with unique markers and colors. ### Components/Axes - **Y-axis**: AUC (0.55 to 1.00, increments of 0.05) - **X-axis**: Number of training bags (0 to 500, increments of 100) - **Legend**: Located in the bottom-right corner, mapping colors/markers to models: - **Blue stars**: MI-SVM - **Orange triangles**: instance+MAX - **Green circles**: embedding+MAX - **Red squares**: Attention - **Purple diamonds**: Gated-Attention ### Detailed Analysis 1. **MI-SVM (Blue stars)**: - Starts at ~0.85 AUC with 0 training bags. - Rapidly increases to 1.00 AUC by 100 bags. - Maintains 1.00 AUC for all subsequent bag counts. - Error bars: Smallest variability (~±0.01). 2. **instance+MAX (Orange triangles)**: - Starts at ~0.60 AUC with 0 bags. - Sharp upward trend, reaching 1.00 AUC by 200 bags. - Error bars: Largest variability (~±0.03 at 100 bags). 3. **embedding+MAX (Green circles)**: - Begins at ~0.95 AUC with 0 bags. - Remains flat at 1.00 AUC after 50 bags. - Error bars: Minimal (~±0.005). 4. **Attention (Red squares)**: - Starts at ~0.75 AUC with 0 bags. - Gradual increase to 1.00 AUC by 300 bags. - Error bars: Moderate (~±0.02 at 100 bags). 5. **Gated-Attention (Purple diamonds)**: - Starts at ~0.90 AUC with 0 bags. - Reaches 1.00 AUC by 150 bags. - Error bars: Small (~±0.01). ### Key Observations - **Early Performance**: MI-SVM and embedding+MAX achieve high AUC with minimal training bags. - **Catch-Up Dynamics**: instance+MAX and Gated-Attention require more bags to reach peak performance. - **Variability**: instance+MAX and Attention exhibit larger error bars, suggesting instability in training. - **Plateaus**: All models except instance+MAX and Attention stabilize at 1.00 AUC after a certain point. ### Interpretation The data demonstrates that model architecture and training strategy significantly impact performance. MI-SVM and embedding+MAX are highly efficient, achieving optimal AUC with fewer training bags. In contrast, instance+MAX and Attention require more data to converge, with instance+MAX showing the greatest variability. Gated-Attention balances early performance and stability. The error bars highlight the importance of training data quantity and quality, with instance+MAX being the most sensitive to fluctuations. This suggests that while some models are robust to limited data, others demand careful tuning and larger datasets to avoid overfitting or underperformance.