GraphFramEx: Towards Systematic Evaluation of Explainability Methods for Graph Neural Networks

TL;DR

GraphFramEx addresses the absence of a standardized evaluation protocol for GNN explainability in node classification by introducing a groundtruth-agnostic framework that blends two fidelity notions into a single characterization score and distinguishes explanations as necessary or sufficient. It formalizes a three-axis design for explanations (focus, mask type, transformation), applies a rigorous evaluation across synthetic and real datasets, and demonstrates that no single explainer dominates across all dimensions, with Saliency often excelling in realism and efficiency while other methods excel in specific aspects. The framework is validated via a real-world eBay fraud case study and culminates in a practical decision tree to select suitable explainers based on user needs, paving the way for more robust, production-aware explainability assessments. The work also provides an online leaderboard and emphasizes the need to move beyond simplistic synthetic benchmarks toward evaluations that reflect real-world graph complexity and model accuracy interactions.

Abstract

As one of the most popular machine learning models today, graph neural networks (GNNs) have attracted intense interest recently, and so does their explainability. Users are increasingly interested in a better understanding of GNN models and their outcomes. Unfortunately, today's evaluation frameworks for GNN explainability often rely on few inadequate synthetic datasets, leading to conclusions of limited scope due to a lack of complexity in the problem instances. As GNN models are deployed to more mission-critical applications, we are in dire need for a common evaluation protocol of explainability methods of GNNs. In this paper, we propose, to our best knowledge, the first systematic evaluation framework for GNN explainability, considering explainability on three different "user needs". We propose a unique metric that combines the fidelity measures and classifies explanations based on their quality of being sufficient or necessary. We scope ourselves to node classification tasks and compare the most representative techniques in the field of input-level explainability for GNNs. For the inadequate but widely used synthetic benchmarks, surprisingly shallow techniques such as personalized PageRank have the best performance for a minimum computation time. But when the graph structure is more complex and nodes have meaningful features, gradient-based methods are the best according to our evaluation criteria. However, none dominates the others on all evaluation dimensions and there is always a trade-off. We further apply our evaluation protocol in a case study for frauds explanation on eBay transaction graphs to reflect the production environment.

Figures (12)

• Figure 1: General protocol. The explanation focus is the phenomenon or the model. (1) A GNN model learns to predict the label $\hat{y}$ of each node in the input graph. For the explanation of node labels (true or predicted), we use this pre-trained model. The explainability method generates a soft mask $M_E$, which operates on the input graph to return a subgraph $G_S$. (2) The goal is to reproduce a target label: $y$ or $\hat{y}$. (3) The mask is transformed to output the final explanatory subgraph $G^t_S$. (4) We evaluate $G^t_S$ by comparing its predicted label to our target.
• Figure 2: Characterization score for $w_+ = w_- = 0.5$
• Figure 3: Results on real datasets. (left) Performance and computation time. (right) Type of explanation returned by each explainability method. sa - Saliency. ig - Integrated Gradient.
• Figure 4: Average performance when explaining only correct (left) or only wrong (right) predictions on 5 real datasets. sa - Saliency. ig - Integrated Gradient.
• Figure 5: GraphFramEx decision tree for a mask transformation $topk=10$.