Interpretability-Guided Bi-objective Optimization: Aligning Accuracy and Explainability
Kasra Fouladi, Hamta Rahmani
TL;DR
This paper introduces Interpretability-Guided Bi-objective Optimization (IGBO), a framework that trains interpretable models by embedding domain knowledge as a DAG prior and using Temporal Integrated Gradients (TIG) for feature attribution. It addresses OOD concerns by learning a Routing Oracle that generates data-manifold-aware integration paths, enabling stable, differentiable interpretability during training. The authors develop a projection-based gradient update to balance accuracy and interpretability, accompanied by convergence and robustness guarantees. An extensive experimental plan on synthetic, clinical, and financial time-series demonstrates the approach’s ability to enforce DAG constraints with minimal accuracy loss and to outperform standard regularization baselines. Overall, IGBO provides a rigorous, scalable path to training models whose internal reasoning aligns with expert-domain structures, with broad implications for safety-critical applications.
Abstract
This paper introduces Interpretability-Guided Bi-objective Optimization (IGBO), a framework that trains interpretable models by incorporating structured domain knowledge via a bi-objective formulation. IGBO encodes feature importance hierarchies as a Directed Acyclic Graph (DAG) and uses Temporal Integrated Gradients (TIG) to measure feature importance. To address the Out-of-Distribution (OOD) problem in TIG computation, we propose an Optimal Path Oracle that learns data-manifold-aware integration paths. Theoretical analysis proves convergence properties and robustness to mini-batch noise, while empirical results on time-series data demonstrate IGBO's effectiveness in enforcing DAG constraints with minimal accuracy loss, outperforming standard regularization baselines.