Targeted data augmentation for improving modelrobustness against systematic bias

Abstract

The paper proposes a new and effective bias mitigation method called Targeted Data Augmentation (TDA). Since removingbiases is a tedious, always difficult and, on the other hand, not necessarily an effective approach the authors propose toskillfully insert them, instead. To show the efficiency and to validate the proposed approach, two representative and verydiverse datasets: the dataset of clinical skin lesions and the dataset of male and female faces, were selected to serve as thebenchmarks. The existing biases were first manually examined, identified, and annotated. Then, the use of CounterfactualBias Insertion, has provided the confirmation that the biases like the frame, ruler, and glasses, strongly affect the models. Tomake the models more robust against them, Targeted Data Augmentation was used: in short, the samples were modifiedduring training by randomly inserting biases. The proposed method resulted in a significant decrease in bias measures, morespecifically, from a two-fold to more than 50-fold improvement after training with TDA, with a negligible increase in the error rate.