Bonsai: Tree-Based Method Preserves High-Dimensional Data Structure Better Than UMAP
bravo_abad · x · 2026-08-24
Daan H. de Groot et al. introduce Bonsai, a Bayesian method that addresses how UMAP and t-SNE distort high-dimensional data structures in visualization.
Key Innovations:
- Instead of forcing data into 2D coordinates, Bonsai represents high-dimensional data as a tree.
- It reconstructs the maximum-likelihood tree connecting points under a probabilistic model, preserving cell-to-cell distances.
- Results: On synthetic benchmarks, Bonsai nearly perfectly reconstructed structures that PCA and UMAP failed to recover. It also identified a previously undescribed subset of natural-killer cells in cord-blood data.
- Implication: The paper notes that dimensionality reduction is not neutral; changing data representation alters which hypotheses become visually plausible.
Published in Nature Biotechnology (2026).
More from Research
- Sleep-time Compute: letting models think ahead during idle time — gerardsans · 2026-08-25
- Study finds authors' self-rankings predict paper citations better than peer review — kchonyc · 2026-08-25
- AI-GENIE automates psychological scale development using LLMs and network analysis — GolinoHudson · 2026-08-25
- ML Framework Evolves Enzymes to Access All Eight Stereoisomers — bravo_abad · 2026-08-25
- Critiquing poor ML papers, recommending classics of science philosophy — fleetwood___ · 2026-08-25
- Artificial Analysis partners with Liquid for mobile benchmarks — JosephJacks_ · 2026-08-25