基于策略梯度的黑盒模型显著性解释方法OA

As a core enabling technology for trustworthy artificial intelligence,research on the interpretability of machine learning models plays a vital role in automated decision-making systems.Among them,saliency-based explanations,which intuitively reveal the basis of model decisions in the form of heatmaps,have become a key research direction in explainable artificial intelligence.However,existing approaches suffer from fundamental limitations:Perturbation-based methods are constrained by manually designed distributional assumptions,making it difficult to balance computational efficiency and generalization ability;gradient-based methods require prior knowledge of the model's internal structure and thus fail to explain black-box models;and linear-approximation-based methods introduce model bias due to excessive local simplification.To address these limitations,we propose Polixplain,a policy-gradient-based saliency explanation method for black-box models.By using the Deletion Score as the core reward signal and incorporating a differentiable relaxation technique into the policy gradient optimization process,Polixplain enables end-to-end optimization toward the faithfulness objective.Image classification tasks are adopted as the experimental testbed,where the high-dimensional feature space and intuitive visual representation provide an ideal environment for multidimensional evaluation of explanation methods.Polixplain outperforms a range of existing explanation methods in terms of both explanation faithfulness and the human interpretability of the generated heatmaps.