Adaptive Protection Boosts Evolutionary Feature Construction

This research addresses a common challenge in evolutionary feature construction for symbolic regression: the loss of valuable features due to genetic operators. While evolutionary methods are powerful for discovering informative feature transformations, existing approaches often lack mechanisms to explicitly protect these crucial components. The paper proposes an adaptive protection mechanism that uses feature importance metrics to selectively safeguard constructed features during the evolutionary process. More important features receive stronger protection, while less important ones can still be modified, allowing for the incorporation of useful building blocks. Experiments on 98 regression benchmark datasets demonstrate that this mechanism consistently improves solution quality compared to baseline approaches. Furthermore, its application to two credit classification datasets shows that the method effectively enhances search effectiveness beyond symbolic regression, indicating its broader applicability in machine learning.