New Framework Generates Novel Graphs While Preserving Structure

The challenge of generating truly novel data, especially in complex structures like graphs, while ensuring it remains consistent with overall patterns, is a significant area of research. This new framework addresses this by introducing an information-theoretic approach to graph novelty generation. The goal is to create data points that are clearly distinct from known examples but still adhere to the fundamental structural characteristics of the dataset. The proposed method involves embedding the input data into a latent space, where its distribution is then modeled using finite mixture models. Novel samples are generated by applying specific conditions for novelty and reliability, which are formulated using the concept of description length. Novelty is enforced by ensuring that generated samples are poorly explained by any existing mixture component, while reliability ensures their minimal impact on the overall mixture structure, guided by the Minimum Description Length (MDL) principle. Theoretical analysis supports the framework, demonstrating that with appropriate threshold settings, the probabilities of misclassifying non-novel or unreliable samples converge to zero at explicit rates. Empirical evaluations on both synthetic and benchmark graph datasets confirm that this method enables principled novelty generation with quantifiable risk, offering a robust approach to creating genuinely new graph structures.