When high accuracy misleads: Stability limits of supervised feature importance in QSAR biodegradation

Key Points

• High predictive accuracy may not reflect real model stability in supervised learning applications, leading to misleading conclusions.
• Feature importance measures are shown to be unstable, impacting the reliability of QSAR biodegradation predictions under varying conditions.
• This analysis explores the role of model stability in supervised learning, particularly in the context of biodegradation.
• The findings highlight the need for careful evaluation of feature importance to ensure valid environmental assessments.