Fractional class-specific weighted random forest optimization

Random forests are a cornerstone of ensemble learning due to their robustness and scalability. However, their potential is constrained by a rigid voting structure. By assigning equal weight to every tree, the standard model overlooks the heterogeneous performance inherent to individual trees across diverse target classes. In this paper, we introduce the Fractional Class-Specific Weighted Random Forest ( f CWRFO). This approach refines tree level confidence weights through fractional order gradient descent. Indeed, it expands the optimization landscape beyond standard integer derivatives which only represent a specific instance of our framework ( α =1). This mechanism introduces a history-dependent fractional scaling of the weight updates, which adaptively modulates the optimization trajectory and improves convergence stability. Comprehensive benchmarks on 40 UCI datasets demonstrate that f CWRFO improves generalization accuracy and stability, outperforming both standard Random Forests and classical integer-based weighting schemes.