ParamBoost: Gradient Boosted Piecewise Cubic Polynomials

Researchers Nicolas Salvadé and Tim Hillel have introduced ParamBoost, a novel type of Generalized Additive Model (GAM) designed to bridge the gap between predictive power and expert interpretability. Unlike black-box models, GAMs are 'glass-box'—their predictive function is fully observable. ParamBoost's innovation lies in using a gradient boosting algorithm to learn shape functions, where each leaf node fits a cubic polynomial. This approach allows the model to capture complex, non-linear relationships in data while remaining fundamentally interpretable.

Empirical results show that the unconstrained version of ParamBoost consistently outperforms other state-of-the-art GAMs across multiple real-world datasets. However, its true power is flexibility: modelers can impose specific constraints derived from domain knowledge, such as ensuring predictions are monotonic with respect to an input (e.g., 'risk always increases with age') or that a function is convex. These constraints for parametric analysis come at only a modest cost to accuracy, allowing the model to be tailored for applications where reasoning and compliance with expert rules are as critical as raw performance.

The model incorporates several key constraints to ensure well-behaved, interpretable functions: continuity up to the second derivative (C²), monotonicity, convexity, feature interaction limits, and broader model specification rules. This makes it particularly valuable for high-stakes fields like finance, medicine, or operational research, where understanding *why* a model makes a prediction is non-negotiable. ParamBoost effectively offers a new toolkit for building AI that is both powerful and accountable.