Graph Similarity Training With Self-Guided Node Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing explainable graph similarity methods rely heavily on node features with rich semantic knowledge and often use toy datasets, leading to misleading inference results due to the soft nature of cross-graph attention, and lack effective methods for explainability inference.
Innovation Solution
A method involving a graph similarity model trained with a multi-task learning framework using 1×1 cross-graph attention and self-guided data augmentation, which enhances explainability by refining node alignments and improving the model's understanding of graph similarity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If cross-graph attention is used to enable information flow between graphs, then the model can capture graph similarity, but the soft nature of attention leads to misleading explainability results
Solution Approach 1:
The patent introduces an intermediary component that decouples the similarity computation from the explainability inference. Instead of directly using soft attention maps for explanation, the system uses the attention mechanism only for computing similarity scores, and introduces a separate discrete alignment inference process that produces reliable explainability results by selecting specific node correspondences rather than relying on continuous attention weights.
Solution Approach 2:
The patent segments the graph similarity task into two distinct components: (1) a similarity scoring function using cross-graph attention, and (2) an explainability inference function that operates on the attention maps to produce discrete node alignments. This segmentation allows each component to optimize for its specific purpose without the conflicting requirements of the other.
2Measurement precision
If existing explainable graph similarity methods are used, then node alignments can be obtained, but they are heavily dependent on node features with rich semantic knowledge and toy datasets
Solution Approach 1:
The patent changes the parameter representation from continuous attention weights to discrete node alignment indicators. By formulating the explainability task as selecting discrete node correspondences rather than interpreting continuous attention values, the method becomes more adaptable to real-world scenarios where clear node-to-node mappings are more meaningful and less dependent on rich semantic features.
3Ease of operation
If cross-graph attention maps are analyzed at inference, then explainability can be achieved, but choosing an effective method for explainability inference remains challenging
Solution Approach 1:
The patent makes the system self-service by using the attention mechanism's own outputs to guide the explainability inference process. The cross-graph attention maps are used to identify candidate node pairs for alignment, and the system automatically selects the most plausible alignments based on attention scores without requiring external intervention or complex post-processing methods.
Data Source
AI summary
A method for training a model for determining graph similarity is disclosed. The method comprises receiving a first graph and a second graph as training inputs, the first graph and the second graph each including nodes connected by edges. The method further comprises applying a model to the first graph and the second graph to determine (i) pairs of aligned nodes between the first graph and the second graph and (ii) a first training loss. The method further comprises generating a first augmented graph by modifying the first graph depending on the pairs of aligned nodes. The method further comprises applying the model to the first graph and the first augmented graph to determine a second training loss. The method further comprises refining the model based on the first training loss and the second training loss.