On Pitfalls of $\textit{RemOve-And-Retrain}$: Data Processing Inequality Perspective

This paper assesses the reliability of the RemOve-And-Retrain (ROAR) protocol, which is used to measure the performance of feature importance estimates. Our findings from the theoretical background and empirical experiments indicate that attributions that possess less information about the decision function can perform better in ROAR benchmarks, conflicting with the original purpose of ROAR. This phenomenon is also observed in the recently proposed variant RemOve-And-Debias (ROAD), and we propose a consistent trend of blurriness bias in ROAR attribution metrics. Our results caution against uncritical reliance on ROAR metrics.