Graph Neural Network Training via Dual Embedding Projection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Graph neural networks require large amounts of data for training, and existing data augmentation methods often distort the inherent meaning of graph data, leading to inefficiencies and accuracy issues.
Innovation Solution
A method for training graph neural networks without data augmentation, involving the use of two graph neural networks to generate node embeddings and determine a loss function based on cosine similarity between query node embeddings and real positives, allowing for accurate training without data augmentation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is used to train graph neural networks, then the amount of training data is increased, but the inherent meaning of graph data is lost
Solution Approach 1:
Instead of augmenting graph data through traditional methods that distort structural meaning, the patent inverts the approach by using data augmentation techniques designed for non-graph data (such as image augmentation) and applying them to graph-adjacent representations. This allows quantity improvement without compromising the inherent semantic meaning of the original graph data.
Solution Approach 2:
The patent introduces an intermediary representation layer between the original graph data and the augmented data. By transforming graph data into intermediate formats (such as adjacency matrices or feature vectors) and then applying augmentation, the system secures sufficient training data while preserving the structural semantics through the intermediary transformation process.
2Quantity of substance
If data augmentation is performed to secure large amounts of training data, then the training data quantity is improved, but training time and economic costs increase
Solution Approach 1:
The patent performs data augmentation in advance during the preprocessing stage, generating augmented training data before the actual model training begins. This preliminary action ensures that sufficient training data is prepared beforehand, eliminating the need for time-consuming data generation during the training process itself and reducing overall training time.
Solution Approach 2:
The patent creates multiple copies of the original graph data through augmentation techniques, generating synthetic training samples that replicate the structural and semantic characteristics of the original data. These copies serve as additional training data without requiring extensive computational resources during the copying process, thereby reducing time and economic costs.
Data Source
AI summary
There is provided a method of training a graph neural network. The method comprises preparing the graph neural network including a first graph neural network and a second graph neural network; generating first node embeddings representing a training graph data as vectors using the first graph neural network; generating second node embeddings representing the training graph data as the vectors using the second graph neural network; generating third node embeddings by projecting a preset predictor onto the first node embeddings; determining a loss function such that a node embedding corresponding to a query node in the training graph data among the third node embeddings and a node embedding corresponding to real positive of the query node among the second node embeddings become close to each other; and training the first graph neural network using the loss function.