Beyond prediction accuracy: Assessing SHAP-driven interpretability for refractory high entropy alloys hardness

This study critically examines feature selection methods in materials informatics, addressing concerns about reliability in supervised machine learning applications. Using an alloy oxidation temperature dataset with 886 instances, we compare supervised models (XGBoost, Random Forest), unsupervised techniques (Feature Agglomeration, Highly Variable Gene Selection), and Spearman correlation for predictive accuracy and feature importance stability. Results reveal that Feature Agglomeration achieves superior performance (R²=0.6345) with perfect feature ranking stability, while supervised models show concerning instability despite reasonable accuracy. XGBoost particularly underperforms (R²=0.3588) with dramatic feature ranking inconsistency. Notably, experimental parameters were consistently identified as crucial by better-performing methods, aligning with metallurgical principles. This work demonstrates that high prediction accuracy doesn't guarantee reliable feature importance identification and recommends using multiple complementary approaches with stability testing for robust materials property-structure relationship discovery, challenging current practices in materials informatics.