Training machine learning algorithms for automatic facial coding: The role of emotional facial expressions’ prototypicality
Fig 2
Effect of training material on overall classification performance.
Note. Overall classification accuracy for the models trained with the standardized dataset (Panel A) and the models trained with the unstandardized dataset (Panel B); for different test sets on the x-axis. Models were evaluated with either a hold-out test set from the same material or with the respective other dataset. Ntest = 147 in both test sets, Nstandardized = 480 in the standardized dataset, and Nunstandardized = 483 in the unstandardized dataset. Error bars represent the 95% confidence interval [59]. The no-information rate is depicted as a dotted line (14.3%). All accuracies were significantly higher than the no information rate (exact binomial test, all ps < .0001, 1-sided).