Explaining Vision GNNs: A Semantic and Visual Analysis of Graph-based Image Classification

TL;DR

This paper tackles explainability for Vision GNNs by introducing a layer-wise semantic analysis of patch-based graphs in Vision GNNs like ViG. It proposes five quantitative metrics ($S^l_{emb}$, $D^l$, $S^l_{vis}$, $p^l$, $Q^l$) and heatmap visualizations to study how graph connections reflect semantic and spatial relationships across layers, and to compare standard vs adversarial inputs. The findings show a progressive shift from local to global feature integration, with deeper layers forming class-specific representations, yet these representations sometimes diverge from human perception. The work provides a framework to interpret Vision GNN decisions and highlights the potential and limitations of graph-based explanations, with implications for robustness and architectural design.

Abstract

Graph Neural Networks (GNNs) have emerged as an efficient alternative to convolutional approaches for vision tasks such as image classification, leveraging patch-based representations instead of raw pixels. These methods construct graphs where image patches serve as nodes, and edges are established based on patch similarity or classification relevance. Despite their efficiency, the explainability of GNN-based vision models remains underexplored, even though graphs are naturally interpretable. In this work, we analyze the semantic consistency of the graphs formed at different layers of GNN-based image classifiers, focusing on how well they preserve object structures and meaningful relationships. A comprehensive analysis is presented by quantifying the extent to which inter-layer graph connections reflect semantic similarity and spatial coherence. Explanations from standard and adversarial settings are also compared to assess whether they reflect the classifiers' robustness. Additionally, we visualize the flow of information across layers through heatmap-based visualization techniques, thereby highlighting the models' explainability. Our findings demonstrate that the decision-making processes of these models can be effectively explained, while also revealing that their reasoning does not necessarily align with human perception, especially in deeper layers.

Figures (3)

• Figure 1: Visual depiction of ViG's representation of a "dog" image at different layers (2, 3). Subgraphs refer to the central green patch and its neighbors (red patches) at that layer. Diagrams represent object-based modularity ($Q$), visual similarity ($S_{vis}$) and spatial distance ($D$) with dots corresponding to the 2nd and 3rd layer.
• Figure 2: Heatmap Visualization of intermediate graphs (layers 4 and 10), for two ImageNet images, and metric evolution across all layers Visual Similarity, Spatial Distance (top image) and Embedding Similarity and Prediction probability (bottom image).
• Figure 3: Heatmap Visualization of intermediate graphs (layers 1, 8 and 15), for intra-domain ImageNet (a) and adversarial ImageNet-a (b) images, and metric evolution across all layers for Embedding Similarity and Graph Modularity (c).