Spurious Correlations in Machine Learning: A Survey

Machine learning systems are known to be sensitive In recent years, spurious correlations have been studied under to spurious correlations between nonessential various names, such as shortcuts, dataset biases, group features of the inputs (e.g., background, robustness, simplicity bias, and so on. We have seen significant texture, and secondary objects) and the corresponding progress in analyzing and mitigating spurious correlations labels. These features and their correlations in various areas such as computer vision (Wang et al., with the labels are known as "spurious" 2021), natural language processing (Du et al., 2022b), and because they tend to change with shifts in realworld healthcare (Huang et al., 2022). Despite the progress, there data distributions, which can negatively impact lacks a survey in this area that formally defines spurious correlations the model's generalization and robustness.