Explainability in Graph Neural Networks: A Taxonomic Survey (2012.15445v3)

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.

• The paper introduces a taxonomy categorizing instance-level and model-level approaches to elucidate GNN decision-making.
• It proposes a unified evaluation framework using metrics like fidelity, sparsity, and stability to benchmark explainability methods.
• Experimental results, including superior performance from SubgraphX, highlight practical gains in GNN transparency.

Explainability in Graph Neural Networks: A Taxonomic Survey

The paper "Explainability in Graph Neural Networks: A Taxonomic Survey" by Hao Yuan, Haiyang Yu, Shurui Gui, and Shuiwang Ji offers a comprehensive survey of current methods aiming to address the challenge of explainability in Graph Neural Networks (GNNs). This survey categorizes various existing approaches, evaluates their methodologies, and presents a unified testbed to facilitate further research and development.

Overview of GNN Explainability

Graph Neural Networks have gained prominence due to their effectiveness in handling graph-structured data across diverse domains such as social networks, molecular biology, and financial modeling. However, like many deep learning models, they often operate as "black boxes" without providing insights into their decision-making processes. The growing demand for transparency in AI systems signals the necessity of developing techniques that can elucidate the functioning of GNNs in a manner comprehensible to humans.

Taxonomy of Explanation Methods

The paper introduces a novel taxonomy of GNN explainability techniques, categorizing them into instance-level and model-level methods.

1. Instance-level Methods: These provide input-dependent explanations for each graph.
   • Gradients/Features-based Methods: These leverage backpropagation to determine the importance of input features, extending traditional methods used in image classification tasks.
   • Perturbation-based Methods: By monitoring changes in output following input perturbations, they identify influential features.
   • Surrogate Methods: These use interpretable models to approximate complex GNN predictions, acting as proxies for providing explanations.
   • Decomposition Methods: These distribute predictions back to input features to provide insights into GNN decision-making.
2. Model-level Methods: These aim to provide global explanations without the need to focus on specific input examples. The primary approach, XGNN, uses graph generation to discover patterns that influence GNN behavior.

Evaluation Framework

To address the absence of standardized benchmarks, the survey introduces a comprehensive evaluation framework comprising datasets and metrics. The framework allows for the assessment of explanation fidelity, sparsity, and stability, which are crucial measures to quantify the effectiveness of explainability methods.

Experimental Analysis

The authors provide an experimental benchmark assessing various methods on multiple datasets, encompassing both synthetic and real-world scenarios. SubgraphX, on account of its ability to consider structural information, consistently demonstrates superior performance. Such results highlight the necessity of methods that embrace graph-specific designs.

Practical Implications and Future Directions

The implications of this survey extend into real-world domains where understanding and trust in GNNs are essential. The standardized testbed facilitates easier replication and comparison of methods, paving the path for new advancements in GNN explainability. The paper also emphasizes the need for more intuitive datasets, such as those derived from text data, which are easier for domain experts to work with without extensive background knowledge.

The research speculates a trend towards developing methods that inherently integrate explainability within the GNN design rather than as a post hoc solution. This approach could foster more transparent models that offer insights organically through their inherent architecture.

Conclusion

By providing a structured taxonomy and a rigorous evaluation framework, this paper significantly contributes to the understanding of GNN explainability. Future research directions may focus on integrating explainability seamlessly within GNN structures and expanding the application of these methods in interdisciplinary scenarios where graph data predominate.