When perfect is a warning: explanations expose shortcuts that accuracy hides in cancer gene-expression classifiers
When the classes in a training set come from different data sources, a classifier can score near-perfectly while learning where a sample came from. We show