Even small correlation and diversity shifts pose dataset-bias issues

Alceu Bissoto, Catarina Barata, Eduardo Valle, Sandra Avila

Abstract: Highlights•Integrated analysis of correlation & diversity shifts with varying intensities.•Finding 1: Low bias training still propagates correlation shifts.•Finding 2: Spurious feature reliance lessens when altered/interfered during testing.•Finding 3: Diversity shift can reduce spurious correlation reliance.•We verify our findings in a real-world bias case, going beyond synthetic sets.