Interpretability and explainability

As graph learning models become increasingly complex and are applied to critical domains such as healthcare, finance, and social sciences, the need for interpretable and explainable models has grown significantly. Here, we explore two key aspects of interpretability and explainability in graph learning.

Explaining GNN decisions

GNNs often act as black boxes, making it challenging to understand why they make certain predictions. This lack of transparency can be problematic in high-stakes applications such as drug discovery or financial fraud detection. To address this, several approaches have been developed to explain GNN decisions:

• One prominent method is GNNExplainer, which identifies important subgraphs and features that influence a model’s predictions. It does this by optimizing a mutual information objective between a conditional distribution of the GNN’s predictions and a simplified explanation.
• Another approach is GraphLIME, an extension of the Local Interpretable Model-agnostic Explanation (LIME) framework for graph-structured data. It explains individual predictions by learning an interpretable model locally around the prediction.
• Gradient-based methods, such as Grad-CAM adapted for graphs, provide explanations by visualizing the gradients of the output with respect to intermediate feature maps, highlighting important regions of the input graph. Some recent works also focus on counterfactual explanations for GNNs, generating minimal changes to the input graph that would alter the model’s prediction.

These approaches help in understanding model decisions and identify potential biases or vulnerabilities in the model.

Visualizing graph embeddings

Graph embeddings, which represent nodes or entire graphs as vectors in a low-dimensional space, are fundamental to many graph learning tasks. However, interpreting these embeddings can be challenging due to their high-dimensional nature. Various techniques have been developed to visualize and understand these embeddings:

• Dimensionality reduction techniques, such as t-distributed Stochastic Neighbor Embedding (t-SNE) or Uniform Manifold Approximation and Projection (UMAP), are commonly used to project high-dimensional embeddings into 2D or 3D spaces for visualization. These methods aim to preserve local relationships between points, allowing for the identification of clusters and patterns in the embedding space.
• Interactive visualization tools, such as TensorBoard Projector or Embedding Projector, allow users to explore embeddings dynamically, zooming in on specific regions and examining relationships between nodes. Some advanced approaches combine embedding visualization with the original graph structure. For instance, GraphTSNE integrates graph structural information into the t-SNE algorithm, producing layouts that reflect both the embedding similarity and the graph topology.
• Another innovative approach is the use of graph generation techniques to visualize embeddings. By training a graph generative model on embeddings and original graphs, one can generate synthetic graphs that represent different regions of the embedding space, providing intuitive visualizations of what the embeddings have captured.

By addressing these aspects of interpretability and explainability, researchers aim to bridge the gap between the performance of complex graph learning models and the need for transparent, trustworthy AI systems. As the field progresses, we can expect to see further integration of these techniques into mainstream graph learning frameworks, making them more accessible to practitioners across various domains. The development of interpretable and explainable graph learning models not only enhances trust and adoption but also opens new avenues for scientific discovery and knowledge extraction from complex graph-structured data.