图神经网络可解释性Slide：GNNExplainer: Generating Explanations for Graph Neural Networks
the initial comprehensive, model-independent methodology designed to offer transparent explanations for the predictions made by any Graph Neural Network (GNN) across diverse graph-based machine learning tasks. When provided with an instance, GNNEXPLAINER systematically identifies a concise subgraph structure and pinpointed node attributes that are essential to the model's predictive outcomes. Furthermore, this innovative framework is capable of generating consistent and succinct explanations applicable to an entire dataset. By framing GNNEXPLAINER as an optimisation problem, we aim to maximise the mutual information between a GNN's predictions and the distribution of potential subgraph configurations.
/
第一种通用的、与模型无关的方法旨在为基于图的机器学习任务上任何基于GNN的模型提供透明度高的预测解释。当输入一个实例时,GNNEXPLAINER能够系统地识别出紧凑的子图结构以及那些对模型预测至关重要的人因属性。此外,该创新方法还能够生成一致且简明扼要的一组适用于整个数据集的解释说明。We frame GNNEXPLainer as an optimisation problem that seeks to maximise the mutual information between a G NN's prediction and the distribution of possible subgraph structures. This approach enables us to identify the most influential components within complex graph data while maintaining computational efficiency.
为了解决什么问题？
Graphs serve as powerful data representations but pose challenges due to the necessity of modeling intricate relational data alongside node-specific feature information [45, 46]. To tackle these challenges effectively, Graph Neural Networks (GNNs) have become prominent solutions for machine learning on graphs. Their capacity lies in recursively integrating information from neighboring nodes within the graph structure while naturally capturing both structural and feature aspects.
为了要做GNN的可解释性？
The capability to comprehend GNN’s predictions holds significant importance and practical utility for multiple reasons: (i) it helps build trust in the model’s reliability, (ii) it enhances transparency within growing high-stakes scenarios concerning fairness, privacy, and other safety challenges [11], and (iii) it enables practitioners to gain insights into network characteristics, identify systematic error patterns before models are deployed in real-world environments.
参考素材：