In this work, we propose a methodology for investigating the application of semantic attention to enhance the explainability of Graph Neural Network (GNN)-based models, introducing semantically-informed perturbations and establishing a correlation between predicted feature-importance weights and model accuracy. Graph Deep Learning (GDL) has emerged as a promising field for tasks like scene interpretation, leveraging flexible graph structures to concisely describe complex features and relationships. As traditional explainability methods used in eXplainable AI (XAI) cannot be directly applied to such structures, graph-specific approaches are introduced. Attention mechanisms have demonstrated their efficacy in estimating the importance of input features in deep learning models and thus have been previously employed to provide feature-based explanations for GNN predictions. Building upon these insights, we extend existing attention-based graph-explainability methods investigating the use of attention weights as importance indicators of semantically sorted feature sets. Through analysing the behaviour of predicted attention-weights distribution in correlation with model accuracy, we gain valuable insights into feature importance with respect to the behaviour of the GNN model. We apply our methodology to a lidar pointcloud estimation model successfully identifying key semantic classes that contribute to enhanced performance effectively generating reliable post-hoc semantic explanations.

翻译：本文提出了一种研究方法，探讨如何利用语义注意力增强基于图神经网络（GNN）模型的可解释性，引入语义信息注入的扰动，并建立预测特征重要性权重与模型准确性之间的关联。图深度学习（GDL）已成为场景理解等任务中的新兴研究方向，借助灵活的图结构，能够简洁地描述复杂特征及关系。由于传统可解释人工智能（XAI）方法无法直接应用于此类结构，研究者提出了专门面向图的方法。注意力机制在深度学习中已被证明能有效估计输入特征的重要性，因此此前已被用于为GNN预测提供基于特征的解释。基于这些见解，我们扩展了现有基于注意力的图解释方法，探究将注意力权重作为按语义排序的特征集的重要性指标。通过分析预测注意力权重分布与模型准确性的关联行为，我们获得了关于GNN模型行为中特征重要性的宝贵见解。我们将该方法应用于激光雷达点云估计模型，成功识别了提升性能的关键语义类别，有效生成了可靠的事后语义解释。