Structural Graph Convolutional Neural Network for Graph Invariant Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Convolutional Neural Networks (CNNs) are not graph-invariant and cannot effectively perform tasks on non-Euclidean data structures, such as graphs, limiting their application to node-level analysis, while graph kernels are insufficient for classifying or clustering whole graphs or subgraphs.
Innovation Solution
A Structured Graph Convolutional Neural Network (SGCNN) is developed, capable of performing graph-invariant learning tasks at the graph and subgraph level, using a graph-based convolutional neural network with a graph feature compression layer, neighbor node aggregation layer, and subgraph convolution layer to generate feature vectors and classify structural relationships.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If spectral domain approaches are used for graph convolution, then convolution operation can be applied on graphs, but the method is not graph invariant and cannot be applied to different graphs
Solution Approach 1:
The patent introduces an intermediary mapping mechanism that transforms graph data into a standardized representation space. This intermediary layer enables the convolutional neural network to process different graphs invariantly by converting their specific structures into a common representation framework, thereby achieving both graph invariance and cross-graph applicability.
2Adaptability or versatility
If vertex domain approaches are used for graph convolution, then graph invariance is achieved, but the method only works at node-level and cannot classify whole graphs or subgraphs
Solution Approach 1:
The patent segments the graph processing task into multiple levels: node-level feature extraction, subgraph-level aggregation, and graph-level classification. By dividing the complex graph analysis into manageable segments, the system achieves graph invariance at each level while extending operational scope to handle both node-level and graph-level tasks without excessive complexity.
Solution Approach 2:
The patent transitions from vertex-domain (node-level) analysis to subgraph-domain analysis by adding an additional layer of abstraction. This dimensional change enables the system to aggregate node features into subgraph representations and perform classification at the graph level, thereby expanding operational scope while maintaining graph invariance.
3Measurement precision
If graph kernels are used to measure similarity between graph pairs, then structure similarity can be captured, but the method is insufficient for classifying or clustering whole graphs or subgraphs with different structures
Solution Approach 1:
The patent replaces traditional graph kernel methods with a convolutional neural network-based approach. This substitution enables the system to not only measure structure similarity but also perform graph and subgraph classification by learning hierarchical representations through convolutional operations, thereby enhancing both measurement precision and classification capability.
Data Source
AI summary
A computer-implemented method for learning structural relationships between nodes of a graph includes generating a knowledge graph comprising nodes representing a system and applying a graph-based convolutional neural network (GCNN) to the knowledge graph to generate feature vectors describing structural relationships between the nodes. The GCNN comprises: (i) a graph feature compression layer configured to learn subgraphs representing embeddings of the nodes of the knowledge graph into a vector space, (ii) a neighbor nodes aggregation layer configured to derive neighbor node feature vectors for each subgraph and aggregate the neighbor node feature vectors with their corresponding subgraphs to yield aggregated subgraphs, and (iii) a subgraph convolution layer configured to generate the feature vectors based on the aggregated subgraphs. Functional groups of components included in the system may then be identified based on the plurality of feature vectors.