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

Generative models often struggle when deployed on target domains that differ from their training (source) distribution. Traditional solutions like fine-tuning are computationally expensive and require domain-specific parameters. This paper asks: can we create a reusable generative framework that adapts to new target domains without parameter updates, using only a small representative set of target data?

The authors propose the Similarity-based Generative Network (SGN). Unlike standard conditional generative models that map labels to fixed embeddings, SGN learns a latent space structured by label-induced pairwise similarities. It uses an encoder-decoder architecture where the latent space is constrained such that the inner product of two samples' latent representations equals one if they share a label and zero otherwise. During the generation phase, the model takes a small labeled representative set from the target domain, encodes these samples, and uses their learned latent representations to guide the generation of new, target-aligned data.

SGN effectively decouples source-domain training from target-domain generation. By enforcing a zero-one similarity matrix in the latent space, the model preserves class semantics while allowing the decoder to synthesize samples that inherit the specific characteristics of the target domain. The authors provide theoretical analysis regarding the required dimensionality of the latent space to satisfy these similarity constraints. Experiments across eleven image and tabular datasets demonstrate that SGN outperforms six state-of-the-art baselines in generating high-quality, class-consistent data under significant distribution shifts.

This work introduces a distribution-agnostic evaluation mechanism for generative models. By moving away from rigid, source-dependent class embeddings toward a flexible, similarity-based latent structure, SGN offers a more efficient and scalable way to perform data augmentation in real-world scenarios where data distributions are constantly evolving.