Skip to main content
. 2018 Nov 22;10:56. doi: 10.1186/s13321-018-0310-y

Fig. 7.

Fig. 7

Comparison of AUC, BEDROC20, and EF1% scores by number of known activities. Illustration of the comparative performance of the algorithms by the number of known affinities per target in the training data. Error bars correspond to standard error. The collaborative filtering based implicit structure methods significantly outperform other algorithms when the training data has 100 or fewer activities. Beyond 100 activities the performance of all algorithms converge