Graph Neural Network Explanation via Orbit Decomposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graph neural networks face challenges in interpreting predictions due to their complex structure, and existing methodologies for providing explanations often lose structural information and cannot offer both global and local explanations simultaneously.
Innovation Solution
The method employs graphlets as basic units for analysis and orbits as explanation components, decomposing the weight matrix into orbit weight and orbit-class score matrices to generate global and local explanations that reflect structural contributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If subgraphs composed of important edges are used as explanation units, then the explanation can be provided for graph neural network predictions, but structural information is lost in the explanation
Solution Approach 1:
The patent segments the graph into graphlets (small induced subgraphs) as basic units, and further decomposes graphlets into orbits (structural patterns). This segmentation allows preservation of structural information while creating manageable explanation units that maintain the original graph's structural characteristics.
Solution Approach 2:
The patent introduces orbits as intermediary concepts that bridge graphlets and global graph structure. Orbits serve as standardized templates that capture structural patterns, allowing the explanation system to reference structural information without directly manipulating complex subgraph configurations.
2Adaptability or versatility
If existing explanation methodologies are used, then local explanations can be provided, but global explanations cannot be provided simultaneously
Solution Approach 1:
The patent creates a unified explanation framework where the same decomposition of weight matrices into orbit weight matrices and orbit-class score matrices serves both local and global explanation purposes. The system can generate explanations at different scopes (local node level or global graph level) using the same underlying mechanism, achieving multi-functionality.
Solution Approach 2:
The patent adds a global dimension to local explanations by introducing orbit-level aggregations. While maintaining node-level detailed explanations, the system simultaneously provides graph-level summaries by aggregating orbit contributions across the entire graph, effectively adding an explanatory dimension rather than choosing between local or global.
3Loss of information
If the weight matrix is decomposed into orbit weight matrix and orbit-class score matrix, then both global and local explanations can be generated, but the decomposition process increases computational complexity
Solution Approach 1:
The weight matrix decomposition is segmented into two distinct components: orbit weight matrices that capture structural pattern weights and orbit-class score matrices that capture classification decision weights. This segmentation makes the decomposition process more manageable and interpretable, reducing the perceived complexity by breaking down a single complex matrix into two conceptually distinct matrices.
Data Source
AI summary
There is provided an apparatus of generating explanations for a graph neural network. The apparatus comprises the graph neural network that embeds an input graph into a node representation matrix H, and then outputs a result matrix Z in which a score of a specific class for each node is represented corresponding to the node representation matrix using a weight matrix W.