SGN: A Similarity-based Generative Network for Data Generation under Distribution Shift
SGN is trained once on labeled source data, then uses a small labeled target set to steer generation without updating its parameters.
The paper frames the problem as target-domain augmentation under source-to-target distribution shift. Its method builds a latent space around label-induced pairwise similarities while keeping reconstruction through an encoder-decoder setup. At generation time, encoded target representatives are combined in that latent space so samples pick up target traits while preserving class consistency. The authors report experiments on image and tabular datasets and include analysis of the similarity structure’s realizability and dimensionality requirements. ArXiv · AI/CL/LG's note
The paper frames the problem as target-domain augmentation under source-to-target distribution shift. Its method builds a latent space around label-induced pairwise similarities while keeping reconstruction through an encoder-decoder setup. At generation time, encoded target representatives are combined in that latent space so samples pick up target traits while preserving class consistency. The authors report experiments on image and tabular datasets and include analysis of the similarity structure’s realizability and dimensionality requirements. ArXiv · AI/CL/LG's note
score 4