Wasserstein Distances Made Explainable: Insights Into Dataset Shifts and Transport Phenomena

Understanding why datasets differ matters for fairness: The paper demonstrates that Wasserstein distances can be broken down to show which specific features or data subgroups contribute most to differences between datasets. This is important for identifying whether dataset shifts (changes in data over time) are driven by meaningful factors or potential biases that could affect AI system performance.

Explainability reveals hidden patterns in data: Rather than just knowing that two datasets are different, this method shows what is driving those differences. The three use cases demonstrate practical applications where understanding these drivers helps practitioners detect problematic changes in their data that simple distance metrics alone would miss.

Responsible AI requires transparency in data analysis: By making Wasserstein distances "explainable," this work addresses a key principle of responsible AI—that technical decisions affecting people should be interpretable and auditable. Organizations can now justify why they believe their training data differs from real-world data, rather than relying on a black-box number.

Detecting dataset inhomogeneities protects against discrimination: The paper's focus on identifying "inhomogeneities within data" (uneven distributions across subgroups) is directly relevant to preventing algorithmic bias. If certain demographic groups are underrepresented or systematically different in training data, this method can help surface those problems before deployment.