SR-tSNE Enhances Embedding Rank Preservation via Differentiable Loss

Smooth Rank t-SNE (SR-tSNE) introduces a differentiable rank term to the standard t-SNE loss using a sigmoid approximation. This modification ensures that high-dimensional neighbor rankings are preserved in low-dimensional embeddings, outperforming 12 baselines across multiple data modalities.