Graph Neural Network Augmentation Using Explainer Subgraphs

Solution Overview

Problem

Existing data augmentation techniques for graph neural networks (GNNs) fail to preserve structural and semantic properties of graphs, leading to out-of-distribution samples and increased sample complexity, which affects the efficiency and accuracy of GNN training.

Innovation Solution

Utilize explanation-assisted empirical risk minimization (EA-ERM) to generate label-preserving perturbations of explainer subgraphs, creating an augmented training dataset that maintains the structural and semantic integrity of graphs, thereby enhancing the generalization capabilities of GNNs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If existing data augmentation techniques are used for GNN training, then the training dataset size is increased, but the structural and semantic properties of graphs are not preserved, leading to out-of-distribution samples

Engineering Contradiction:

Improvetraining dataset sizeVSAvoidstructural and semantic property preservation

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

The method performs preliminary identification of explanation subgraphs that capture critical structural and semantic properties before generating augmentations. By pre-identifying these important subgraphs and using them to guide the augmentation process through EA-ERM, the system ensures that generated samples preserve essential graph characteristics while increasing dataset size.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

Explanation subgraphs serve as intermediaries between the original graph data and the augmented training samples. These subgraphs capture the essential structural and semantic properties and are used as guides during the augmentation process, ensuring that generated samples maintain fidelity to the original data distribution while increasing dataset diversity.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Quantity of substance

If data augmentation is applied to increase training data, then sample complexity increases, but the accuracy and efficiency of GNN training deteriorates due to out-of-distribution samples

Engineering Contradiction:

Improvetraining data volumeVSAvoidtraining efficiency and accuracy

Core Design Contradiction:

Quantity of substanceVSProductivity

Solution Approach 1:

The system employs explanation-assisted empirical risk minimization (EA-ERM) that uses feedback from the GNN's own explanations to guide the augmentation process. The GNN generates explanations for its predictions, and these explanations are used to create targeted augmentations that are more likely to improve performance. This feedback loop ensures that augmentations are meaningful and contribute to training efficiency rather than introducing noise.

Inventive Principle:

Principle #23Feedback

3Quantity of substance

If conventional data augmentation methods are used, then the training dataset is enlarged, but the generalization capabilities of GNNs are not improved due to loss of graph integrity

Engineering Contradiction:

Improvedataset sizeVSAvoidgeneralization capability

Core Design Contradiction:

Quantity of substanceVSAdaptability or versatility

Solution Approach 1:

The method performs preliminary identification of explanation subgraphs that capture critical structural and semantic properties before generating augmentations. By pre-identifying these important subgraphs and using them to guide the augmentation process through EA-ERM, the system ensures that generated samples preserve essential graph characteristics while increasing dataset size.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

Explanation subgraphs serve as intermediaries between the original graph data and the augmented training samples. These subgraphs capture the essential structural and semantic properties and are used as guides during the augmentation process, ensuring that generated samples maintain fidelity to the original data distribution while increasing dataset diversity.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS20250348756A1Explanation-assisted data augmentation for graph neural network training

Publication Date: 2025.11.13 NEC LABORATORIES AMERICA INC

AI summary

Systems and methods for explanation-assisted data augmentation for training graph neural networks. The GNN can be trained using a training dataset to generate explainer subgraphs of labeled graphs. The explainer subgraphs can be transformed into perturbed subgraphs by utilizing explanation assisted empirical risk minimization (EA-ERM) learned by the GNN to generate an augmented training dataset. The GNN can be further trained with the augmented training dataset and the training dataset to perform downstream tasks using input data with corresponding labels.