Mitigating class imbalance in the SAFARI benchmark

Determine how to mitigate the imbalance issue arising from SAFARI’s empirical distribution of real-world industrial HARA cases, including its uneven representation of risk categories.

Background

SAFARI preserves the empirical distribution of industrial HARA records rather than balancing the dataset artificially. Because real safety-engineering practice contains rare but safety-critical situations, the resulting benchmark has imbalanced risk-category distributions. The paper explicitly identifies mitigation of this imbalance as unresolved future development, while noting that the imbalance does not undermine the benchmark’s purpose of reflecting real-world industrial conditions.