跳转至

忠实、充分且可解释：通过离散扩散反演重新审视图反事实解释

文章背景与核心概要

图神经网络（GNN）在化学、生物学和网络分析等领域的图结构数据上取得了极高的预测性能，但它们缺乏内在的可解释性，这阻碍了其在高风险和安全关键场景中的应用。反事实解释（Counterfactual Explanations）通过识别改变模型预测所需的最小结构修改，为这一问题提供了解决方案。然而，由于图数据具有离散、组合式的搜索空间以及特定领域的约束（如分子图中的化学化合价），在图上生成有效的反事实极其困难。现有的方法通常被迫妥协：要么编辑偏离了数据流形，要么搜索无法覆盖完整的编辑空间。

为了克服这些局限，作者提出了基于反演的图扩散反事实解释（Graph Diffusion Counterfactual Explanation via Inversion, 简称 GDCE-I）。通过将离散去噪扩散模型与新颖的离散反演方案相结合，GDCE-I 能够在利用整个领域编辑空间的同时进行分布感知（distribution-aware）的编辑。此外，该论文引入了一个基于既定解释诉求（desiderata）的统一评估框架，用于在统一协议下对图反事实方法进行一致的基准测试。在四个基准测试中，GDCE-I 显著优于相关工作，同时能够为分子图生成具可解释性的分布内（in-distribution）解。

Faithful, Sufficient and Understandable: Rethinking Graph Counterfactual Explanations via Discrete Diffusion Inversion

Authors: David Bechtoldt, Sidney Bender
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv ID: 2608.12083
Submitted: August 12, 2026

Summary

Graph Neural Networks (GNNs) achieve strong predictive performance across domains like chemistry, biology, and network analysis, but lack intrinsic explainability—hindering their adoption in high-stakes and safety-critical settings. Counterfactual explanations address this by identifying minimal structural modifications required to alter a model's prediction.

However, generating valid counterfactuals on graphs is challenging due to discrete, combinatorial search spaces and domain-specific constraints (e.g., chemical valency in molecular graphs). Existing methods typically force a compromise: either edits stray off the data manifold, or the search fails to span the full edit space.

To overcome this, the authors propose Graph Diffusion Counterfactual Explanation via Inversion (GDCE-I). By combining a discrete denoising diffusion model with a novel discrete inversion scheme, GDCE-I enables distribution-aware edits across the entire domain edit space. Furthermore, the paper introduces a unified evaluation framework based on defined explanation desiderata to consistently benchmark graph counterfactual methods. Across four benchmarks, GDCE-I significantly outperforms related work while generating interpretable, in-distribution solutions for molecular graphs.

摘要

图神经网络（GNN）在化学、生物学和网络分析等领域的图结构数据上取得了强大的预测性能，但它们没有提供对其预测的内在解释。这限制了它们在高风险和安全关键环境中的采用。反事实解释通过揭示能够改变模型预测的最小结构修改来解决这个问题。然而，在图上生成这样的修改是困难的。搜索空间是离散且组合的，并且有效的答案必须遵守分类节点和边缘类型，以及诸如分子图情况下的化学化合价等领域规则。现有的解释器放弃了以下两点之一：要么编辑没有保持在数据流形上，要么搜索没有跨越完整的编辑空间。我们提出了基于反演的图扩散反事实解释（GDCE-I），它两点都不放弃。带有新型离散反演方案的离散去噪扩散模型能够利用整个领域编辑空间进行分布感知的编辑。我们进一步通过推导解释诉求框架并将其应用于同一共享协议下的每个方法，解决了图反事实评估不完整和不一致的问题。在四个基准测试中，GDCE-I 在所定义的框架上大幅优于相关工作。对于分子领域，我们进一步定性地表明 GDCE-I 获得了可解释的分布内解。

Graph Neural Networks (GNNs) achieve strong predictive performance on graph-structured data across domains such as chemistry, biology, and network analysis, yet they provide no intrinsic explanation of their predictions. This limits their adoption in high-stakes and safety-critical settings. Counterfactual explanations address this by revealing the minimal structural modifications that would change a model's prediction. On graphs, however, such a modification is hard to produce. The search space is discrete and combinatorial, and a valid answer must respect categorical node and edge types together with domain rules such as chemical valency in the case of molecular graphs. Existing explainers give up one of two things. Either edits are not held on the data manifold, or the search does not span the full edit space. We propose Graph Diffusion Counterfactual Explanation via Inversion (GDCE-I), which gives up neither. A discrete denoising diffusion model with a novel discrete inversion scheme enables distribution-aware edits leveraging the whole domain edit space. We further address the incomplete and inconsistent evaluation of graph counterfactuals by deriving a framework of explanation desiderata and applying it to every method under one shared protocol. Across four benchmarks, GDCE-I outperforms related work by a large margin on the defined framework. For the molecular domain, we further qualitatively show that GDCE-I attains interpretable in-distribution solutions.

链接与资源