LiBaGS: Lightweight Boundary Gap Synthesis for Targeted Synthetic Data Selection

LiBaGS is an efficient, generator-independent approach for the selective generation of synthetic training data. It evaluates potential synthetic samples by integrating factors such as proximity to decision boundaries, predictive uncertainty, density of real data, and validity of support, ensuring that the chosen samples are both informative and likely to reside within the real data manifold. The method employs a boundary-gap allocation strategy that focuses on sparse yet realistic decision-boundary areas, rather than merely increasing data volume or choosing only the most uncertain samples. Additionally, LiBaGS identifies when sufficient synthetic samples have been incorporated through a marginal-value stopping criterion, applies softer labels near ambiguous boundaries, and incorporates a diversity objective to prevent redundant selections. Experiments indicate that LiBaGS enhances model performance by effectively addressing gaps in the training distribution.