Visualizing Graph Neural Networks with CorGIE: Corresponding a Graph to Its Embedding

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 two GNN experts. Availability: Open-source code, supplemental materials & video at https://osf.io/tr3sb/.

VA + Embeddings STAR: A State‐of‐the‐Art Report on the Use of Embeddings in Visual Analytics

“…Most of the papers in this set focus on mining graphs and/or sequential data. The embeddings may represent the nodes of a network [CZIM18, XXM19, XTYL20, PCZ*21, LTHL21, SDXR22, LWBM22] or sequences of events [LDL*20,XTYL20], ultimately aiding in identifying structural, temporal, and multivariate properties within groups of nodes [PCZ * 21], communities [CZC * 17], ensembles [FFST19, WJM * 22] (see Figure 6(c)), etc.…”

“…Examples of node‐link diagrams for trees and graphs/networks, mainly or partly derived using embedding data. (a) CorGIE by Liu et al [LWBM22] for visualizing graph neural networks. (b) BiaScope by Rissaki et al [RSL * 22] for interactive investigation of unfairness for graph embeddings.…”

“…Neural Network Interpretation We observe less direct user involvement for general neural network visualization papers in the corpus. Like neural language models, generalized neural network‐related VA tools in our corpus focus on model understanding and validation [BPP * 16, SGR * 20, FZCM20, CDHP21, RSL * 22, PDD * 22,LWBM22,SGL22]. Users may explore decision and feature spaces for counterfactual reasoning [SGL22], kernel actions [BPP * 16], and neuron group interactions [PDD * 22], or validate unfairness and errors in AI systems [CDHP21,RSL * 22].…”

“…Correspondingly, the predominant visual representations are 2‐dimensional scatter plots projecting the embedding space to give an overview of the dataset. One particularly interesting approach in this set is from Liu et al [LWBM22] (see Figure 7(a)). They present node‐link diagrams to visualize node embeddings for graph neural networks.…”

BiaScope: Visual Unfairness Diagnosis for Graph Embeddings

“…The explanations have the form of small visualizable graph motifs and important node features. On the other hand, the CorGIE tool [41] focuses on the correspondence between the graph topology and the node embeddings. It utilizes a k-hop graph layout to show topological neighbors in hops and their clustering structure.…”