Summary
- Researchers proposed Smooth Rank t-SNE (SR-tSNE), a manifold learning method that adds a differentiable rank term to the standard t-SNE loss to preserve internal neighbor order.
- The method uses a sigmoid rank approximation to explicitly force high-dimensional true neighbors to outrank unrelated intruders.
- An extension of the Degree of Locality Preservation metric named Ordinal DLP (O-DLP) was introduced to credit a neighbor only when it returns at its original rank.
- Across eight benchmarks, SR-tSNE outperformed 12 baselines in rank preservation, neighborhood overlap, image label-transfer, single-cell discovery, and intraday trading tasks.
Why it matters
Improving neighborhood order preservation in dimensionality reduction techniques like t-SNE directly enhances downstream machine learning applications such as clustering, classification, and financial forecasting.