Abstract
Dimensionality reduction has become an essential tool across scientific disciplines, yet the field remains fragmented. This review establishes a unified taxonomy of methods, provides best practices for researchers, and benchmarks popular algorithms to clarify the inherent trade-offs between preserving local and global data structures. By framing embeddings as epistemic tools for hypothesis generation, the authors offer a roadmap for moving beyond "black box" applications toward more rigorous and interpretable data exploration.
Introduction
In fields ranging from single-cell biology to the digital humanities, researchers are increasingly confronted with high-dimensional data—datasets where each sample is characterized by thousands of features. To extract meaning from this complexity, "embeddings" are used to project data into 2D or 3D visualizations. As shown in the citation trends below, the popularity of these methods, particularly neighbor-embedding techniques like t-SNE and UMAP, has exploded over the last decade.
However, this rapid adoption has led to a fragmented landscape where tools are often treated as "black boxes," leading to over-interpretation of artifacts or the loss of critical global structures. This paper addresses the lack of clear, cross-disciplinary guidance on method selection and validation.
A Taxonomy of Embedding Methods
The authors categorize the vast array of embedding techniques into a clear taxonomy based on their underlying mathematical principles and optimization goals.
- Linear Methods (e.g., PCA): These remain the "gold standard" for preserving global structure and providing high interpretability.
- Distance-Preserving (e.g., MDS): These focus on maintaining the actual distances between all pairs of points.
- Spectral Methods (e.g., Laplacian Eigenmaps, PHATE): These leverage graph theory to identify underlying manifolds or continuous trajectories.
- Neighbor-Embedding (e.g., t-SNE, UMAP): Currently the most popular for identifying discrete clusters by prioritizing local neighborhood preservation.
- Parametric Methods (e.g., Autoencoders): These utilize neural networks to learn a reusable mapping function.
Key Insights: The Local-Global Trade-off
A central theme of the review is the inherent tension in dimensionality reduction: the local-global trade-off. Methods optimized for showing clusters (t-SNE/UMAP) often distort the "big picture" distances between those clusters. Conversely, methods that preserve the global structure (PCA/MDS) often fail to resolve fine-grained local clusters.
The authors introduce a quantitative evaluation framework to measure this, using local quality (nearest neighbor preservation) and global quality (correlation between high- and low-dimensional pairwise distances). Their benchmarks reveal that while t-SNE consistently leads in local quality, PCA and MDS are superior for global structure.
Benchmarking and Best Practices
The paper evaluates six major methods across diverse real-world datasets, including Wikipedia text, mouse cortex cells, primate brain organoids, and human genotypes from the 1000 Genomes Project. Key findings include:
- Contextual Superiority: PCA was superior for the 1000 Genomes data due to its continuous geographic variation, while UMAP and t-SNE were essential for separating the distinct semantic topics in Wikipedia text.
- Continuous vs. Discrete: For developmental data (organoids), spectral methods like PHATE better captured the continuous "flow" of time, whereas t-SNE tended to fragment trajectories into artificial clumps.
- Initialization: For non-convex methods like t-SNE, starting the optimization with a PCA-based layout significantly improves the preservation of global structure.
Limitations and Future Outlook
Despite the power of these tools, significant challenges remain. Most non-linear embeddings are "black boxes," lacking clear interpretability regarding which original features drive a point's position. Furthermore, the field lacks "absolute" quality measures to determine if an embedding is "good enough" for a specific scientific conclusion.
Future work should focus on improving scalability for multi-million point datasets, developing "human-in-the-loop" interactive tools, and creating methods that explicitly explain the feature contributions to the low-dimensional projection.
Summary updated