A Feedback-Enhanced Graph Generative Model With Tunable Features

Graphs are data structures that represent abstract relationships between elements and are widely used across various domains. The ability to generate graphs with specific features is crucial for many applications. Recent advances in deep learning have led to progress in graph generation, enabling models to reproduce structural features observed in real-world graphs. A significant recent focus in this area is conditional graph generation, which allows models to tune desired features during the generation process. Among recent efforts in this direction, GraphTune was developed to support continuous tuning of graph features. Building on this approach, this paper introduces F-GraphTune, a model designed to improve the accuracy of feature specification in generated graphs. F-GraphTune incorporates a Feature Estimator (FE) into the GraphTune framework, which provides feedback on the feature values of the generated graphs. Experiments using a real-world graph dataset demonstrate that F-GraphTune outperforms existing models, including GraphTune, with up to a 53.7% improvement in accuracy of feature specification.