Back to AI Research

AI Research

Towards Faithful Graph Explanations with Synergisti... | AI Research

Key Takeaways

  • Towards Faithful Graph Explanations with Synergistic Edge Effects via Granular Balls Graph Neural Networks (GNNs) are powerful tools for analyzing complex da...
  • Instance-level explanations aim to reveal the rationale behind a model's decisions for a specific graph.
  • Previous methods explain graph neural networks (GNNs) by selecting important edges to induce subgraphs, where edge importance is assessed by perturbing each edge and observing changes in the model predictions.
  • However, they often neglect the synergistic effects among edges, which are crucial for accurately characterizing edge importance.
  • To address this issue, we propose SeeExplainer, a parameter-free explainer to interpret GNNs.
Paper AbstractExpand

Instance-level explanations aim to reveal the rationale behind a model's decisions for a specific graph. Previous methods explain graph neural networks (GNNs) by selecting important edges to induce subgraphs, where edge importance is assessed by perturbing each edge and observing changes in the model predictions. However, they often neglect the synergistic effects among edges, which are crucial for accurately characterizing edge importance. To address this issue, we propose SeeExplainer, a parameter-free explainer to interpret GNNs. Specifically, we first introduce a granular-ball graph refinement mechanism that decomposes a graph into several disjoint granular-balls with no fixed size, and utilize them as nodes to construct a structural graph. This process can better capture the synergistic effects among edges. Then, we perturb nodes and edges in the structural graph to generate explanatory subgraphs based on their respective contributions. Experiments on several graph classification datasets of different networks show that SeeExplainer outperforms state-of-the-art baselines.

Towards Faithful Graph Explanations with Synergistic Edge Effects via Granular Balls
Graph Neural Networks (GNNs) are powerful tools for analyzing complex data in fields like chemistry and social networks, but their "black-box" nature makes it difficult to understand why they make specific decisions. While existing explanation methods often focus on identifying important individual edges, they frequently overlook the "synergistic effects"—the idea that certain groups of edges work together to form meaningful structures that are more important than the sum of their parts. This paper introduces SeeExplainer, a new method designed to interpret GNNs by capturing these synergistic relationships through a process called granular-ball computing.

Capturing Synergistic Effects

The core challenge in graph explanation is that edges do not always act independently. For example, in chemical molecules, specific functional groups consist of multiple edges that only carry scientific meaning when viewed as a whole. SeeExplainer addresses this by decomposing a graph into "granular-balls"—disjoint sub-structures of varying sizes. By treating these granular-balls as nodes in a new, simplified "structural graph," the method preserves the original relationships while allowing the model to analyze the graph at a more meaningful, coarse-grained level.

How SeeExplainer Works

The process begins by covering the entire graph with an initial granular-ball. The system then uses a two-step splitting strategy: 1. Coarse-splitting: The graph is divided into several initial granular-balls based on node degree, simulating the human cognitive concept of "global precedence." 2. Fine-splitting: Each granular-ball is iteratively refined into smaller pieces based on a quality metric. If splitting a ball into two smaller ones improves the overall structural quality, the process continues until it converges.
Once the structural graph is constructed, SeeExplainer evaluates the importance of these granular-balls and their connections by measuring how much the GNN’s prediction changes when they are removed.

Generating Explanations

After calculating the contribution of each structural unit and edge, SeeExplainer generates an explanatory subgraph. It uses the average contribution of the entire graph as a threshold; any structural unit or edge that exceeds this average is considered significant. By aggregating these high-impact components, the method produces a clear, localized subgraph that highlights the rationale behind the GNN's decision.

Performance and Impact

SeeExplainer is a parameter-free approach, meaning it does not require complex tuning to function. By shifting the focus from individual edges to these synergistic structural units, the method provides a more faithful interpretation of GNN behavior. Experiments on various real-world graph classification datasets demonstrate that SeeExplainer outperforms current state-of-the-art baseline methods, offering a more effective way to enhance the transparency and reliability of graph-based artificial intelligence.

Comments (0)

No comments yet

Be the first to share your thoughts!