[论文解读] Interpretable Prototype-based Graph Information Bottleneck

Sang‐Woo Seo, Sungwon Kim+1 more|arXiv (Cornell University)|Oct 30, 2023

Explainable Artificial Intelligence (XAI)被引用 8

一句话总结

PGIB 将原型学习与图信息瓶颈结合，以在 GNN 预测中识别关键子图，从而提高准确性和可解释性。

ABSTRACT

The success of Graph Neural Networks (GNNs) has led to a need for understanding their decision-making process and providing explanations for their predictions, which has given rise to explainable AI (XAI) that offers transparent explanations for black-box models. Recently, the use of prototypes has successfully improved the explainability of models by learning prototypes to imply training graphs that affect the prediction. However, these approaches tend to provide prototypes with excessive information from the entire graph, leading to the exclusion of key substructures or the inclusion of irrelevant substructures, which can limit both the interpretability and the performance of the model in downstream tasks. In this work, we propose a novel framework of explainable GNNs, called interpretable Prototype-based Graph Information Bottleneck (PGIB) that incorporates prototype learning within the information bottleneck framework to provide prototypes with the key subgraph from the input graph that is important for the model prediction. This is the first work that incorporates prototype learning into the process of identifying the key subgraphs that have a critical impact on the prediction performance. Extensive experiments, including qualitative analysis, demonstrate that PGIB outperforms state-of-the-art methods in terms of both prediction performance and explainability.

研究动机与目标

• 为解释性 GNNs 做动机，识别图中的核心子结构。
• 将原型学习整合到图信息瓶颈以捕捉关键子图。
• 实现与预测子结构对齐的可解释原型。
• 证明在分类性能和可解释性方面优于最先进方法。

提出的方法

• 将 GIB 目标重新表述以纳入原型和子图级信息。
• 使用子图提取层通过随机遮掩来最小化互信息 I(G; G_sub)。
• 引入一个带有固定数量的类别级原型和基于相似度的评分机制的原型层。
• 优化包含分类、IB 项和连通性正则化的综合损失。
• 合并相似原型并将原型投射到训练子图以实现可解释性。

实验结果

研究问题

• RQ1 如何将原型整合到图信息瓶颈中以聚焦关键子图？
• RQ2 相较于基线，PGIB 是否在预测准确性和解释性方面均有所提升？
• RQ3 哪些机制（遮掩、原型相似性、合并）能提升子图和原型的鲁棒性与可解释性？

主要发现

数据集	GCN	GIN	GAT	ProtGNN	GIB	VGIB	GSAT	PGIB	PGIB_cont
MUTAG	74.50±7.89	80.50±7.89	73.50±7.43	80.50±9.07	79.00±6.24	81.00±6.63	80.88±8.94	85.00±7.07	85.50±5.22
PROTEINS	72.83±4.23	70.30±4.84	71.35±4.85	73.83±4.22	75.25±5.92	73.66±3.32	69.64±4.71	77.14±2.19	77.50±2.42
NCI1	73.16±3.49	75.04±2.08	66.05±1.03	74.13±2.10	64.65±6.78	63.75±3.37	68.13±2.64	77.65±2.20	78.25±2.13
DD	72.53±4.51	72.04±3.62	70.81±4.33	69.15±4.33	72.61±8.26	72.77±5.63	71.93±2.74	73.36±1.80	73.70±2.14

• PGIB 在四个图分类数据集（MUTAG、PROTEINS、NCI1、DD）上优于最先进的基于原型和 IB 的方法。
• 对比变体 PGIB_cont 往往比变分 IB 版本获得更高的准确性。
• 原型合并可减少模型复杂性并在不牺牲性能的前提下提升可解释性。
• 原型投影将每个原型与最近的训练子图绑定，提升解释的可解释性。
• 保真度分析显示基于 PGIB 的解释更好地捕捉了重要的预测成分，相较基线方法。