Figure 3

The predictive performance versus number of positive examples and the position of index keywords. This figure illustrates how the performance of BICEPP is related to the number of positive examples (A) and position of index keywords in the respective feature ranks (B). Each data point represents the best AUC (out of the 4 machine learning algorithms studied in this paper) performed on one of the 484 drug characteristics listed in Table 1. As illustrated in the shaded area in Figure 3(A), the predictive performance of BICEPP had a higher variability in datasets with less than 10 positive examples. The boxed area (*) in Figure 3(B) represents a list of "surprising characteristics", whose predictive powers were high but the index keywords were not discriminative. The contents are listed in more detail in Table 3. Refer to the main text for details.