DT-PBO: an Interpretable Tree-based Surrogate Model for Preferential Bayesian Optimization

Researchers introduce DT-PBO, a tree-based surrogate model for Preferential Bayesian Optimization that prioritizes interpretability over traditional Gaussian Process approaches. The method achieves competitive performance on benchmark functions while providing transparent insights into decision-maker preferences, addressing critical needs in high-stakes domains like healthcare.

DT-PBO represents a meaningful shift in how optimization problems involving human preferences can be solved with transparency. Traditional Preferential Bayesian Optimization relies on Gaussian Process surrogates that excel at capturing complex preference patterns but operate as black boxes, creating friction in sectors where stakeholders must understand why a particular solution was selected. This research bridges that gap by developing interpretable decision trees trained directly from pairwise comparison data, supplemented with Laplace approximation for probabilistic uncertainty quantification.

The broader context reflects growing tension between model performance and explainability across AI and machine learning. As algorithmic systems increasingly influence critical decisions in medicine, finance, and policy, regulators and practitioners demand transparency. The shift toward interpretable-by-design models aligns with this trajectory, though historically at a performance cost.

DT-PBO's competitive convergence with GP-based approaches on benchmark functions, combined with superior performance on rugged optimization landscapes, suggests the method has practical value beyond regulatory compliance. The fast computational runtime and demonstrated robustness to noise indicate potential deployment advantages in resource-constrained environments.

Market implications extend to organizations investing in AI decision-support systems where explainability is non-negotiable. Healthcare institutions, financial compliance teams, and insurers face mounting pressure to document decision rationales. Tools enabling preference learning with transparency could capture significant value in these domains. Looking ahead, watch whether this approach generalizes to higher-dimensional optimization problems and whether adoption accelerates as regulatory frameworks increasingly mandate AI transparency.

• →DT-PBO combines interpretable decision trees with Bayesian optimization to model human preferences transparently.
• →The method achieves convergence competitive with opaque Gaussian Process surrogates across benchmark tests.
• →Shallow decision trees enable direct visualization of how decision-maker preferences relate to solution features.
• →Superior performance on rugged landscapes suggests advantages over existing approaches on complex preference problems.
• →Fast computation and noise robustness indicate practical viability for real-world high-stakes applications.