Figure 4

NN training and testing. In the figure we highlight these two processes with a different colour code, red for the training and blue for the testing. We followed a two-fold heterogeneous cross-validation scheme [50] in which the original dataset was split in two (training and test sets). A resampling protocol was applied to correct for class-imbalance effects [51], resulting in 100 training sets with the same proportion of correct and incorrect observations. Each training set was then utilised to train a NN. We applied the latter to the events in the test set and computed the success rate. The success rate given in the article is the average of the success rates for the 200 NN.