Abstract
Graph neural networks (GNNs) are a class of powerful machine learning tools that model node relations for making predictions of nodes or links. GNN developers rely on quantitative metrics of the predictions to evaluate a GNN, but similar to many other neural networks, it is difficult for them to understand if the GNN truly learns characteristics of a graph as expected. We propose an approach to corresponding an input graph to its node embedding (aka latent space), a common component of GNNs that is later used for prediction. We abstract the data and tasks, and develop an interactive multi-view interface called CorGIE to instantiate the abstraction. As the key function in CorGIE, we propose the K-hop graph layout to show topological neighbors in hops and their clustering structure. To evaluate the functionality and usability of CorGIE, we present how to use CorGIE in two usage scenarios, and conduct a case study with five GNN experts. Availability: Open-source code at https://github.com/zipengliu/corgie-ui/, supplemental materials & video at https://osf.io/tr3sb/.
Introduction
Graph Neural Networks (GNNs) have become the standard for modeling relational data, excelling in tasks like link prediction and node classification. Despite their success, GNNs remain "black boxes" for many developers. Traditional evaluation relies heavily on global quantitative metrics (e.g., F1 scores) or simple visualizations of the high-dimensional latent space. However, these methods fail to explain why a model makes certain predictions or if it is truly capturing the underlying graph structure.
The fundamental challenge is a lack of correspondence. Developers need to know if the proximity of nodes in the latent space reflects meaningful patterns in the original graph's topology or features. For instance, if two nodes are close in the embedding but share no common neighbors or features, the model may be over-relying on noise. CorGIE addresses this by directly linking the input graph to its learned node embeddings, allowing researchers to move beyond simple accuracy scores.
The CorGIE Approach
CorGIE (<u>Cor</u>responding a <u>G</u>raph to <u>I</u>ts <u>E</u>mbedding) is a "grey-box" visualization tool that focuses on the relationship between three key spaces: the Topology space, the Feature space, and the Latent space.
The core technical innovation is the K-hop graph layout. Unlike standard layouts that often result in unreadable "hairballs," the K-hop layout organizes nodes into concentric "boxes" based on their distance (hops) from a selected focal node. This design mirrors the neighborhood aggregation process inherent to GNN architectures, aligning the visualization with the developer's mental model. Within each hop-box, CorGIE applies dimensionality reduction (like UMAP) to reveal how nodes cluster based on their attributes or latent positions, enabling multi-view coordination to trace where a node's topological neighbors ended up in the latent space.
Evaluation and Key Insights
The authors evaluated CorGIE using diverse datasets, including a Movie-User recommendation graph, the Cora citation network, and a large-scale Coauthor dataset. Through expert studies with five GNN researchers, several key insights emerged:
- Identifying "Cold Start" Issues: In recommendation scenarios, CorGIE highlighted that the GNN struggled with users who had very few interactions. Their latent positions appeared random because the model lacked sufficient topological context to learn from.
- Feature Ineffectiveness: The tool allowed experts to see when specific node features failed to contribute to cluster separation in the latent space, suggesting opportunities for model simplification or feature engineering.
- Boundary Node Inspection: Researchers used CorGIE to investigate "boundary nodes"—data points situated between topic clusters—to understand why the model misclassified them, often finding that these nodes shared characteristics of multiple categories.
Conclusion and Future Work
CorGIE demonstrates that interactive visualization can bridge the gap between abstract GNN embeddings and the underlying graph data. While currently optimized for graphs up to 20,000 nodes, future work aims to improve scalability through GPU acceleration and extend support to dynamic graphs and heterogeneous edge types. By making the "hops" of a GNN visible, CorGIE empowers developers to move beyond simple accuracy scores toward a deeper, structural understanding of their models.
Summary updated