Common-Sense Bias Modeling for Classification Tasks
This addresses the need for domain practitioners to understand and mitigate biases from common-sense correlations in real-world datasets, though it is incremental as it builds on existing de-biasing approaches.
The authors tackled the problem of model bias from spurious feature correlations in image datasets by proposing a framework to extract comprehensive biases using textual descriptions, which uncovered novel biases and, when mitigated via data re-weighting, outperformed state-of-the-art unsupervised methods.
Machine learning model bias can arise from dataset composition: correlated sensitive features can distort the downstream classification model's decision boundary and lead to performance differences along these features. Existing de-biasing works tackle the most prominent bias features, such as colors of digits or background of animals. However, real-world datasets often include a large number of feature correlations that intrinsically manifest in the data as common sense information. Such spurious visual cues can further reduce model robustness. Thus, domain practitioners desire a comprehensive understanding of correlations and the flexibility to address relevant biases. To this end, we propose a novel framework to extract comprehensive biases in image datasets based on textual descriptions, a common sense-rich modality. Specifically, features are constructed by clustering noun phrase embeddings with similar semantics. The presence of each feature across the dataset is inferred, and their co-occurrence statistics are measured, with spurious correlations optionally examined by a human-in-the-loop module. Downstream experiments show that our method uncovers novel model biases in multiple image benchmark datasets. Furthermore, the discovered bias can be mitigated by simple data re-weighting to de-correlate the features, outperforming state-of-the-art unsupervised bias mitigation methods.