Zoomed Figure

Dataset LCIT-HC

We further amplified the figures into the very small (2,4,8e-5). We can find out that when FPR increases, the TPR (True positive rate) increases even faster for the RF-CNNensemble. Considering we have pretty much pairs, thus even if the TPR is small, the top predicted results are more likely to be potentially related.

Fig 1. The amplified figure of ROC curves with different ensemble policies.
The dataset is LCIT predicting HC. The x-axis is the FPR, which is cut to 2e-5. The y-axis is the TPR, which is cut to 0.01. From the figure, for the RF-CNN ensemble, FPR increases, and the TPR (True positive rate) increases even faster when FPR is small. Considering we have pretty much pairs, thus even if the TPR is small, the top predicted results are more likely to be potentially related.
Fig 2. The amplified figure of ROC curves with different ensemble policies.
The dataset is LCIT predicting HC. The x-axis is the FPR, which is cut to 4e-5. The y-axis is the TPR, which is cut to 0.01. From the figure, for the RF-CNN ensemble, FPR increases, and the TPR (True positive rate) increases even faster when FPR is small. Considering we have pretty much pairs, thus even if the TPR is small, the top predicted results are more likely to be potentially related.

Fig 3. The amplified figure of ROC curves with different ensemble policies. The dataset is LCIT predicting HC. The x-axis is the FPR, which is cut to 8e-5. The y-axis is the TPR, which is cut to 0.01. From the figure, for the RF-CNN ensemble, FPR increases, and the TPR (True positive rate) increases even faster when FPR is small. Considering we have pretty much pairs, thus even if the TPR is small, the top predicted results are more likely to be potentially related.

Fig 4. The amplified figure of ROC curves with different ensemble policies. The dataset is LCIT predicting HC. The x-axis is the FPR, which is cut to 2e-4. The y-axis is the TPR, which is cut to 0.01. From the figure, for the RF-CNN ensemble, FPR increases, and the TPR (True positive rate) increases even faster when FPR is small. Considering we have pretty much pairs, thus even if the TPR is small, the top predicted results are more likely to be potentially related.

Dataset IT-HC

For other dataset, such

Fig 5. The amplified figure of ROC curves with different ensemble policies. The dataset is IT predicting HC. The x-axis is the FPR, which is cut to 5e-5. The y-axis is the TPR, which is cut to 0.008. From the figure, we could observe the same trend as LCIT-HC task.