[Paper Review] GNNExplainer: Generating Explanations for Graph Neural Networks

Rex Ying, Dylan Bourgeois+3 more|arXiv (Cornell University)|Mar 10, 2019

Explainable Artificial Intelligence (XAI)61 references692 citations

TL;DR

Introduces GNNExplainer, a general, model-agnostic method to produce compact subgraphs and feature subsets that explain GNN predictions, with optimization maximizing mutual information between predictions and subgraph distributions.

ABSTRACT

Graph Neural Networks (GNNs) are a powerful tool for machine learning on graphs.GNNs combine node feature information with the graph structure by recursively passing neural messages along edges of the input graph. However, incorporating both graph structure and feature information leads to complex models, and explaining predictions made by GNNs remains unsolved. Here we propose GNNExplainer, the first general, model-agnostic approach for providing interpretable explanations for predictions of any GNN-based model on any graph-based machine learning task. Given an instance, GNNExplainer identifies a compact subgraph structure and a small subset of node features that have a crucial role in GNN's prediction. Further, GNNExplainer can generate consistent and concise explanations for an entire class of instances. We formulate GNNExplainer as an optimization task that maximizes the mutual information between a GNN's prediction and distribution of possible subgraph structures. Experiments on synthetic and real-world graphs show that our approach can identify important graph structures as well as node features, and outperforms baselines by 17.1% on average. GNNExplainer provides a variety of benefits, from the ability to visualize semantically relevant structures to interpretability, to giving insights into errors of faulty GNNs.

Motivation & Objective

• Motivate the need for interpretable explanations in Graph Neural Networks that combine structure and features.
• Propose a general, model-agnostic framework to generate explanations for any GNN-based model and graph-based task.
• Identify compact subgraphs and relevant node features that drive individual predictions and entire classes of instances.
• Demonstrate the approach on synthetic and real-world graphs and compare against baselines.
• Show that explanations aid visualization, interpretation, and debugging of GNNs.

Proposed method

• Formulate GNN explanations as an optimization problem.
• Maximize mutual information between a GNN’s prediction and the distribution of possible subgraph structures.
• Select a compact subgraph and a small subset of node features that preserve predictive information.
• Extend to generating consistent explanations for an entire class of instances.
• Evaluate on synthetic and real-world graphs against baselines and quantify explanation accuracy (up to 43.0% improvement).

Experimental results

Research questions

• RQ1 How can we derive compact subgraphs that preserve a GNN’s predictive decisions for a given instance?
• RQ2 Which node features are crucial for a GNN’s prediction on a given graph?
• RQ3 Can we extend explanations to be consistent across an entire class of instances?
• RQ4 How does GNNExplainer perform relative to baseline explanation methods on synthetic and real-world graphs?

Key findings

• GNNExplainer identifies important subgraph structures and node features that drive predictions.
• The method outperforms alternative baselines by up to 43.0% in explanation accuracy.
• Explanations provide visualization of semantically relevant structures and offer interpretability and debugging insights for faulty GNNs.