[Paper Review] Explainability in Graph Neural Networks: A Taxonomic Survey

Hao Yuan, Haiyang Yu+2 more|arXiv (Cornell University)|Dec 31, 2020

Advanced Graph Neural NetworksComputer Science94 citations

TL;DR

This survey provides a unified taxonomy of GNN explainability methods, introduces a standardized testbed, and conducts experiments to compare various instance-level and model-level explanations for graph neural networks.

ABSTRACT

Deep learning methods are achieving ever-increasing performance on many artificial intelligence tasks. A major limitation of deep models is that they are not amenable to interpretability. This limitation can be circumvented by developing post hoc techniques to explain the predictions, giving rise to the area of explainability. Recently, explainability of deep models on images and texts has achieved significant progress. In the area of graph data, graph neural networks (GNNs) and their explainability are experiencing rapid developments. However, there is neither a unified treatment of GNN explainability methods, nor a standard benchmark and testbed for evaluations. In this survey, we provide a unified and taxonomic view of current GNN explainability methods. Our unified and taxonomic treatments of this subject shed lights on the commonalities and differences of existing methods and set the stage for further methodological developments. To facilitate evaluations, we generate a set of benchmark graph datasets specifically for GNN explainability. We summarize current datasets and metrics for evaluating GNN explainability. Altogether, this work provides a unified methodological treatment of GNN explainability and a standardized testbed for evaluations.

Motivation & Objective

• Provide a comprehensive, unified review of GNN explanation techniques.
• Propose a taxonomic framework classifying instance-level and model-level explanations.
• Introduce datasets, evaluation metrics, and an open-source library for GNN explainability evaluations.
• Conduct experiments to compare and analyze the performance of multiple explainability approaches.

Proposed method

• Develop a two-branch taxonomy: instance-level and model-level explanations for GNNs.
• Within instance-level explanations, categorize methods into gradients/features-based, perturbation-based, decomposition, and surrogate approaches.
• Survey model-level explanations, currently represented by XGNN, focusing on graph generation for explanations.
• Propose a standardized testbed including datasets, common algorithms, and evaluation metrics for GNN explainability.
• Provide a critical analysis of each method’s methodology, advantages, drawbacks, and differences.

Experimental results

Research questions

• RQ1 What are the main categories and subcategories of GNN explainability methods?
• RQ2 How do instance-level and model-level explanation techniques differ in their goals and outputs?
• RQ3 What datasets and evaluation metrics are suitable for assessing GNN explanations, and what are their limitations?
• RQ4 How do different GNN explanation methods compare in practice across tasks like node, edge, and graph-level predictions?

Key findings

• A unified taxonomy clarifies distinctions among gradients/features-based, perturbation-based, decomposition, and surrogate explanation methods for instance-level explanations.
• Model-level explanation is less explored, with XGNN being a primary example at the time of the survey.
• The paper provides a public testbed including datasets, algorithms, and evaluation metrics to standardize evaluations of GNN explainability methods.
• An open-source library and comprehensive experiments are introduced to compare various explanation techniques and guide future methodological development.