Noise injection proposed to correct categorical bias in tree-based variable importance

Tree-based models such as random forests tend to assign higher importance scores to continuous predictors than to categorical ones. The authors analyze the source of this bias theoretically and suggest a straightforward fix: injecting a small amount of noise into each categorical predictor. The work targets mixed data settings where such scoring imbalance is common.