Invertible Graph Latent Conversion for Reliable Graph Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing graph data generation models face challenges in efficiently generating and verifying the correctness of graph data, particularly in methods like VAE and GAN, where reproducibility and reliability are low, and in one-shot methods, the cost of learning and generating graphs with many nodes is high.
Innovation Solution
A training device and method that utilizes first and second converters to convert graph data into latent values and back, using NVP mapping and neural networks like GNN and GCN, optimizing the conversion process with SGD to achieve high accuracy and efficient generation of graphs with a large number of nodes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If one-shot method is used to generate graph data, then generation speed is improved, but learning cost and computational resource requirements increase significantly for graphs with many nodes
Solution Approach 1:
The patent segments the graph generation process into multiple incremental steps, adding nodes one by one or in batches rather than generating all nodes simultaneously. This divides the complex one-shot generation into manageable stages, reducing the computational burden and memory requirements at each step while maintaining overall generation efficiency.
Solution Approach 2:
The patent employs dynamic graph generation where the graph structure evolves incrementally through multiple processing steps. Nodes and edges are added dynamically based on previous states, allowing the system to adapt computational resources to the current graph size rather than allocating resources for the maximum possible graph size from the start.
2Adaptability or versatility
If VAE or GAN models are used for graph data generation, then generation capability is improved, but reproducibility and reliability of generated data decrease
Solution Approach 1:
The patent incorporates verification mechanisms that check the correctness of generated graph data against expected properties and constraints. This feedback loop ensures that generated graphs meet reliability criteria, allowing the system to maintain high adaptability while improving reproducibility by filtering out invalid generations.
Solution Approach 2:
The patent performs preliminary validation and constraint checking during the generation process rather than only after complete generation. By verifying intermediate results and ensuring compliance with graph properties at each step, the system maintains reliability while preserving the flexibility of VAE/GAN models.
3Reliability
If incremental method is used to add nodes one by one, then reliability of graph structure is improved, but generation time and processing steps increase
Solution Approach 1:
The patent applies partial incremental action by adding nodes in optimized batches rather than strictly one by one. This approach maintains the reliability benefits of incremental construction while reducing total processing time by processing multiple nodes in parallel when possible, balancing structure correctness with generation speed.
Data Source
AI summary
A training device comprises one or more memories and one or more processors. The one or more processors are configured to train a first converter that converts a first feature amount regarding a node of a graph and a second feature amount regarding a structure of the graph into a first latent value through conversion capable of defining inverse conversion, and a second converter that converts the second feature amount into a second latent value through conversion capable of defining inverse conversion, based on the first latent value and the second latent value.