SGN: A Similarity-based Generative Network for Data Generation under Distribution Shift

The paper introduces SGN, a reusable generative framework for augmentation under source-to-target distribution shifts. SGN is trained once on labeled source data and does not update its parameters for new target domains. Its encoder-decoder learns a latent space organized by label-induced pairwise similarities while retaining reconstructive information. At generation time, a small labeled representative set from the target domain is encoded and combined in latent space, enabling generated samples to reflect target-domain characteristics while preserving class consistency. The authors also analyze realizability and dimensionality requirements for the similarity structure, and report experiments on image and tabular datasets.