Core Concepts
This paper introduces robust counterfactual witnesses (RCWs) as explanation structures for graph neural networks (GNNs) that are both factual (preserving the GNN's classification result) and counterfactual (flipping the result if removed), while also remaining stable under a bounded number of graph disturbances.
Abstract
The paper addresses the need for intuitive, robust, and both factual and counterfactual explanation structures for GNN-based node classification tasks. It introduces a new class of explanation structures called robust counterfactual witnesses (RCWs), which satisfy the following properties:
Factual: The RCW subgraph preserves the GNN's classification result for a test node.
Counterfactual: Removing the RCW subgraph from the graph flips the GNN's classification result for the test node.
Robust: The RCW subgraph remains factual and counterfactual even after a bounded number of edge disturbances in the graph.
The paper analyzes the computational complexity of verifying and generating RCWs, establishing hardness results from tractable cases to co-NP hardness. It presents efficient algorithms to verify and generate RCWs, including a parallel algorithm for large graphs. The proposed methods are experimentally validated on real-world datasets, showing their ability to produce intuitive and robust explanations for GNN-based node classification tasks.
The key insights and contributions of the paper are:
Formalization of robust counterfactual witnesses (RCWs) as a novel class of explanation structures for GNNs.
Complexity analysis of RCW verification and generation problems.
Efficient algorithms for verifying and generating RCWs, including a parallel algorithm for large graphs.
Experimental validation demonstrating the effectiveness of RCWs as explanations for GNN-based node classification.
Stats
The paper does not provide any specific numerical data or statistics. It focuses on the theoretical analysis and algorithmic development for generating robust counterfactual explanations for GNNs.
Quotes
The paper does not contain any direct quotes that are particularly striking or support the key arguments.