Figure 8

Test accuracy increases to a model-specific limit as number of training profiles increases. Cross-validation was used to estimate the generalization ability of the global models with different sized training sets. Panels from left to right specify the annotations that were used to train, and panels from top to bottom specify the annotations that were used to test. Note that the results change very little even when training on one data set and testing on another. For each training set size t, the profiles were partitioned into training sets of approximately size t, then were evaluated using the annotations from all the other profiles. Results on these data indicate increasing accuracy (lines) and decreasing standard deviation (shaded bands) as the training set increases. The accuracy of each model quickly attains its maximum, after only about t=10 profiles. In Figure 9, we show the results for all algorithms when trained on t=10 profiles from the detailed data set, and tested on the other profiles in the detailed data set (vertical black line).