Fig. 5
From: SimVec: predicting polypharmacy side effects for new drugs
Negative sampling models. a In terms of validation ROC AUC, Stay Positive model performs significantly better then others. Other models show similar accuracy, albeit Strong NSCAching, NSCaching 6:1 and Expolre NSCaching work slightly better then NSCaching 1:1, Bernoulli and Uniform sampling. b Scores of negative examples, sampled from weak and strong tail caches, are indistinguishable. That is, sampling from strong neighbour cache does not introduce meaningful improvement