[Paper Review] GNNVis: A Visual Analytics Approach for Prediction Error Diagnosis of Graph Neural Networks.

Zhihua Jin, Yong Wang+4 more|arXiv (Cornell University)|Nov 22, 2020

Data Visualization and Analytics60 references4 citations

TL;DR

GNNVis is an interactive visual analytics tool that diagnoses prediction errors in Graph Neural Networks (GNNs) by comparing predictions across GNN, MLP, and GNNWUF models. It uses Parallel Sets, Projection, Graph, and Feature Matrix views to reveal error patterns and disentangle the influence of graph structure versus node features, enabling developers to identify and understand GNN errors systematically.

ABSTRACT

Graph Neural Networks (GNNs) aim to extend deep learning techniques to graph data and have achieved significant progress in graph analysis tasks (e.g., node classification) in recent years. However, similar to other deep neural networks like Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), GNNs behave like a black box with their details hidden from model developers and users. It is therefore difficult to diagnose possible errors of GNNs. Despite many visual analytics studies being done on CNNs and RNNs, little research has addressed the challenges for GNNs. This paper fills the research gap with an interactive visual analysis tool, GNNVis, to assist model developers and users in understanding and analyzing GNNs. Specifically, Parallel Sets View and Projection View enable users to quickly identify and validate error patterns in the set of wrong predictions; Graph View and Feature Matrix View offer a detailed analysis of individual nodes to assist users in forming hypotheses about the error patterns. Since GNNs jointly model the graph structure and the node features, we reveal the relative influences of the two types of information by comparing the predictions of three models: GNN, Multi-Layer Perceptron (MLP), and GNN Without Using Features (GNNWUF). Two case studies and interviews with domain experts demonstrate the effectiveness of GNNVis in facilitating the understanding of GNN models and their errors.

Motivation & Objective

• To address the lack of visual analytics tools for diagnosing prediction errors in Graph Neural Networks (GNNs), which remain largely black-box models.
• To enable model developers and users to identify, validate, and understand error patterns in GNN predictions through interactive visualization.
• To disentangle the relative contributions of graph structure and node features to GNN prediction errors by comparing predictions across three models: GNN, MLP, and GNNWUF.
• To support hypothesis formation about error causes through detailed node-level analysis using Graph and Feature Matrix views.

Proposed method

• Employing Parallel Sets View and Projection View to detect and validate global error patterns across wrong predictions.
• Using Graph View and Feature Matrix View to conduct detailed, node-level analysis for forming hypotheses about error causes.
• Comparing predictions from three models—GNN, MLP, and GNNWUF—to isolate the influence of graph structure versus node features on prediction errors.
• Integrating interactive visual components that allow users to explore relationships between node features, graph topology, and misclassification outcomes.
• Designing a visualization pipeline that supports iterative exploration and hypothesis refinement during GNN error diagnosis.
• Enabling comparative analysis by training and contrasting GNN, MLP (structure-agnostic), and GNNWUF (feature-agnostic) models on the same data.

Experimental results

Research questions

• RQ1 How can visual analytics techniques be effectively applied to diagnose and understand prediction errors in Graph Neural Networks?
• RQ2 What role do graph structure and node features play individually in contributing to GNN prediction errors?
• RQ3 How can interactive visual interfaces support model developers in identifying and validating error patterns across large sets of misclassified nodes?
• RQ4 To what extent can comparing predictions from GNN, MLP, and GNNWUF models help disentangle structural from feature-based error sources?
• RQ5 How effective is the proposed visual analytics tool in supporting expert users during real-world GNN error diagnosis?

Key findings

• The comparison between GNN, MLP, and GNNWUF models revealed that both graph structure and node features significantly influence GNN prediction errors, with distinct error patterns emerging in each model.
• Parallel Sets and Projection Views enabled users to efficiently identify recurring error patterns across misclassified nodes, improving error detection speed and accuracy.
• Graph and Feature Matrix Views allowed users to investigate individual nodes in detail, leading to actionable hypotheses about error causes such as feature sparsity or structural ambiguity.
• Case studies demonstrated that GNNVis helped users uncover previously undetected error patterns related to node feature distribution and local graph topology.
• Expert interviews confirmed that GNNVis enhanced users’ ability to interpret GNN behavior and diagnose errors more effectively than relying on model outputs alone.
• The tool’s interactive design supported iterative exploration, enabling users to refine hypotheses and validate error explanations through visual correlation of features, structure, and predictions.