DREAMS: Preserving both Local and Global Structure in Dimensionality Reduction

Noël Kury [email protected], Dmitry Kobak [email protected], Sebastian Damrich [email protected] · 2025

Read DREAMS: Preserving both Local and Global Structure in Dimensionality Reduction by Noël Kury [email protected], Dmitry Kobak [email protected], Sebastian Damrich [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-13747.txt.

Introduction

Noël Kury, Dmitry Kobak, and Sebastian Damrich blend two map-making philosophies for high-dimensional data, in a 2025 paper. Analysts flatten data into two-dimensional pictures to see structure. Favorite tools keep either the local neighborhoods honest or the global layout, never both. Their method, DREAMS, adds a simple pull toward the global PCA picture inside the local t-SNE picture, yielding a dial between the two extremes.

Eleven real data sets benchmark the blend. Across scales, the embeddings kept both neighborhood and continent structure better than predecessors. One tuning knob trades the microscope for the atlas. The authors offer it as a default first look at messy data.

Catalog record

CREATORNoël Kury [email protected], Dmitry Kobak [email protected], Sebastian Damrich [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-13747

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

DREAMS: Preserving both Local and Global Structure in Dimensionality Reduction Noël Kury [email protected], Dmitry Kobak [email protected], Sebastian Damrich [email protected] Abstract Dimensionality reduction techniques are widely used for visualizing high-dimensional data in two dimensions. Existing methods are typically designed to preserve either local (e.g. t t -SNE, UMAP) or global (e.g. MDS, PCA) structure of the data, but none of the established methods can represent both aspects well. In this paper, we present DREAMS (Dimensionality Reduction Enhanced Across Multiple Scales), a method that combines the local structure preservation of t t -SNE with the global structure preservation of PCA via a simple regularization term. Our approach generates a spectrum of embeddings between the locally well-structured t t -SNE embedding and the globally well-structured PCA embedding, efficiently balancing both local and global structure preservation. We benchmark DREAMS across seven real-world datasets, including five from single-cell transcriptomics and one from population genetics, showcasing qualitatively and quantitatively its superior ability to preserve structure across multiple scales compared to previous approaches. 1 Introduction Real-world data often exhibits high-dimensional and complex structures, making their effective and interpretable visualization a crucial step in exploratory data analysis. Dimensionality reduction methods serve this purpose by projecting high-dimensional data into more interpretable low-dimensional representations while preserving meaningful structures. Among dimensionality reduction methods, principal component analysis (PCA, Hotelling 1933 ) and t -distributed stochastic neighbor embedding ( t -SNE, van der Maaten and Hinton 2008 ) have emerged as two widely used methods, each with distinct strengths. PCA excels at capturing global structures by projecting data onto a lower-dimensional subspace in directions that maximize variance, providing a broad overview of the dataset’s structure. In contrast, t -SNE is a neighbor-embedding method with the objective to map data points that are nearby in high-dimensional space close to one another in the low-dimensional embedding. This focus on neighbor preservation makes t -SNE particularly effective in preserving local structures. When applied individually, both methods suffer from limitations: PCA often overlooks fine local relationships, while t -SNE distorts global structures in favor of local neighborhood preservation. However, in many real-world datasets, e.g. single-cell transcriptomic datasets, both local and global structure is meaningful. Local relationships can reveal microscopic data patterns such as small clusters and cell types. For example, the Tasic et al. 2018 dataset features 122 fine clusters that represent cell types, which are clearly separated in the t -SNE plot ( Figure 1 right). Conversely, the global structure reflects macroscopic data patterns such as separations between broad cell classes ( Tasic et al. 2018 ) or developmental trajectories ( Kanton et al. 2019 ) across an entire cell population. In the Tasic et al. 2018 dataset, this level of structure separates individual cells into non-neurons, excitatory, and inhibitory neurons. This global structure is not evident in the t -SNE plot, but is prominent in the PCA visualization, which in turn fails to separate the finer clusters ( Figure 1 left). Since both local and global structure carry essential and often complementary information, neglecting either structure scale can result in incomplete or biased interpretations, highlighting the importance of preserving both scales simultaneously. Our proposed method DREAMS combines the interpretability and global structure preservation of PCA with the local sensitivity of t -SNE in a simple, yet effective way. On the Tasic et al. 2018 data, it maintains PCA’s global arrangement by non-neurons, inhibitory, and excitatory neurons, while also revealing the finer cluster structure within each broad group, similar to t -SNE ( Figure 1 middle). DREAMS integrates global structure preservation into the t -SNE objective by a PCA-based regularization term applied throughout the entire optimization process. Varying the regularization strength allows DREAMS to transition from the locally well-structured t -SNE embedding to the globally well-structured PCA embedding. Along this spectrum it trades off local and global structure more favorably than its competitors resulting in improved qualitative and quantitative structure preservation, particularly on datasets with hierarchical organization that exhibit both prominent local and global patterns. We chose PCA as the default global method in DREAMS due to its simplicity, inherent interpretability (PCA is linear) and fast runtime. That said, DREAMS also offers regularizing with other global embeddings such as multidimensional scaling (MDS). Figure 1: PCA, DREAMS, and t -SNE embeddings of the Tasic et al. dataset ( Tasic et al. 2018 ) illustrate how DREAMS preserves the global organization seen in the PCA embedding — such as the separation of non-neurons, inhibitory, and excitatory neurons — while also capturing the local cell-type structure that is present in the t -SNE embedding. In summary, we introduce the new method DREAMS making the following contributions: 1. suggest a simple but effective regularization strategy to combine t -SNE’s local with PCA’s global quality; 2. present a spectrum of visualizations from t -SNE to PCA with almost no local/global structure trade-off; 3. perform a benchmark of ten algorithms on seven real-world datasets; 4. provide an open-source implementation of DREAMS based on the openTSNE library. Our code to reproduce the results of this paper is available at https://github.com/berenslab/dreams-experiments . We used a modified openTSNE implementation available at https://github.com/berenslab/DREAMS and a modified CNE implementation available at https://github.com/berenslab/DREAMS-CNE . 2 Related Work The most prominent and well-established visualization methods either excel at preserving local structure, e.g., neighbor-embedding methods, such as t -SNE ( van der Maaten and Hinton 2008 ) and UMAP ( McInnes et al. 2018 ) , or global structure such as PCA ( Pearson 1901 ; Hotelling 1933 ) and MDS ( Kruskal 1964 ) , but not both simultaneously. Therefore, many recent efforts have aimed at producing visualizations with faithful local and global structure. Most of them attempt to improve the global structure preservation of neighbor embeddings. One way to make neighbor embeddings more sensitive to the global structure of a dataset is to widen the set of points that are deemed similar beyond a small set of nearest neighbors. For instance, increasing the perplexity parameter in t -SNE effectively increases the number of considered nearest neighbors but consequently leads to increased runtime ( Kobak and Berens 2019 ; Lee et al. 2015 ; De Bodt et al. 2020 ) . Similarly, several modifications of UMAP, e.g. PaCMAP ( Wang et al. 2021 ) and TriMAP ( Amid and Warmuth 2019 ) , do not only use nearest neighbors, but also consider more distant points in their optimization. In addition to attraction between nearest neighbors, PaCMAP employs weak attraction on mid-near points, while TriMap tries to also preserve the order of similarities in random triplets. Another triplet-based method, ivis ( Szubert et al. 2019 ) , strives to balance local and global structure with a parametric encoder and a margin loss on triplets of embedding distances. A different strategy to improve the global structure in neighbor embeddings is to initialize them with a global embedding, which can improve the global structure of the final embedding, despite no further steps for preserving global structure during optimization ( Kobak and Linderman 2021 ; Wang et al. 2021 ) . A prominent way to trade off local and global structures within the neighbor-embedding framework is the attraction-repulsion spectrum ( Böhm et al. 2022 ; Damrich et al. 2023 ) . Along this

Read the complete text