dtour: a steerable tour de vis through high-dimensional data

Abstract: Understanding high-dimensional data requires projecting it into lower-dimensional spaces, but any single projection inevitably loses information or introduces distortions. Tours address this limitation through animation of 2D projection sequences, yet existing tools present tradeoffs in the freedom and steerability of projection traversal, providing little to no ability to move between expert-guided paths and unrestrained exploration. We present dtour, a tour interface that combines static projection previews, reversible scrubbing along continuous geodesic projection paths, manual projection manipulation, and a wandering grand tour, all within a single progressive exploration interface. dtour scales to millions of points via GPU-accelerated rendering, runs in any modern browser, and integrates with both Python and JavaScript ecosystems. We demonstrate dtour on text, image, and single-cell data for two usage scenarios: gradually revealing structure in high-dimensional data and validating non-linear dimensionality reduction outputs.

• The paper introduces dtour, a unified interface that combines keyframe galleries, guided scrubbing, manual axis manipulation, and grand tours for progressive exploration of high-dimensional data.
• The system interpolates orthonormal projection bases with Catmull–Rom splines and Grassmannian distances, while GPU-accelerated browser rendering reaches about 60 FPS for 5 million points and 25 FPS for 20 million.
• The paper shows that touring across projections, embedding models, and hyperparameters can expose misleading clusters and model-specific behavior, although sequential-tour intermediate frames lack direct latent-space interpretation and user benefits remain unevaluated.

The paper presents dtour, a tour-based interface for exploring high-dimensional data through smoothly interpolated sequences of 2D projections. The central design claim is that existing approaches to multi-projection visualization—scatter plot matrices, animated tours, and manual tours—occupy fixed points on a spectrum of traversal freedom versus steerability, and that these trade-offs can be reconciled only by an interface allowing fluid movement across the spectrum itself. dtour unifies a keyframe preview gallery, a scrubbed guided tour, a manual tour with draggable dimension axes, and a grand tour within a single progressive exploration interface, backed by GPU-accelerated rendering that scales to millions of points in a browser.

Motivation and related work

The authors situate the tool against two bodies of work. First, dimensionality reduction (DR): linear methods such as PCA preserve global structure but hide orthogonal structure, while neighbor embeddings such as t-SNE and UMAP introduce distortions that can misrepresent cluster structure—a concern raised prominently by critiques of single-cell embeddings (Damrich et al., 2022, Singh et al., 2022). Second, tour methods: the grand tour [asimov1985grand], guided tours via projection pursuit [cook1995grand], and manual tours [cook1997manual], formalized mathematically by Buja et al. [buja2005computational]. The authors observe that the dominant tour software ecosystem (tourr, GGobi, spinifex, detourr, langevitour) requires analysts to commit upfront to a single tour mode, and that embedding-visualization tools (Embedding Projector, WizMap, Embedding Atlas, Emblaze) present static projections without smooth, steerable traversal through projection space. dtour's contribution is the unification of these modes with million-point scalability.

Interface design

The interface follows Shneiderman's visual information-seeking mantra with three coordinated components. A keyframe gallery surrounds the central scatter with previews of each projection, including feature-loading text, acting as both a lookahead and orientation device. A guided tour is controlled by a circular slider whose ring-segment widths encode geodesic distances between keyframes—thin segments indicate stretched regions of projection space—supporting scrubbing, scrolling, and playback along a closed loop. A manual tour renders dimension axes as draggable handles whose direction and length encode each variable's contribution to the projection basis; dragging re-orthonormalizes the remaining basis, and a Shift-drag rotates about the residual principal component. Lasso and label-based selection support a select-then-explore workflow across modes. All transitions are smoothly interpolated to preserve object constancy.

Interpolation method

Tour frames are $p \times 2$ orthonormal basis matrices. Distances between bases are computed as the geodesic on the Grassmannian of 2D subspaces, derived from principal angles obtained via SVD of $\mathbf{F}_a^\top \mathbf{F}_z$ (computed analytically for the $2\times 2$ case). Interpolation uses a Catmull-Rom spline applied element-wise followed by Gram-Schmidt orthonormalization. The authors argue this avoids the velocity discontinuities of piecewise geodesic interpolation while guaranteeing orthonormality, and the spline passes exactly through each keyframe with $C^1$-continuous tangents. A precomputed cumulative arc-length table (eight samples per segment) with binary-search lookup yields $O(\log n)$ mapping to arc-length parameter, producing constant angular velocity during scrubbing.

Tour strategies

dtour accepts any sequence of orthonormal bases, organized into two families. Hyperdimensional tours traverse a single high-dimensional space: the little tour cycles through successive PC pairs, and the "le tour" uses Laplacian Eigenmaps with cumulative circular basis construction. Sequential embedding tours interpolate between aligned 2D embeddings of the same points—constructed by warm-started DR runs with Procrustes alignment—with the explicit caveat that intermediate frames are geometrically valid but should not be interpreted as views of latent structure. A notable special case is the attraction-repulsion tour, which sweeps the exaggeration hyperparameter to traverse the spectrum from Laplacian Eigenmaps through UMAP to t-SNE [boehm2022attraction].

Scalability

The implementation is a TypeScript renderer, React component, or Anywidget, with rendering offloaded to a WebGPU/WebGL worker via OffscreenCanvas and a data worker streaming Parquet columns to the GPU. On an Apple M1 Max, the system sustains roughly 60 FPS at up to 5M points, 40 FPS at 10M, and 25 FPS at 20M points. This is a strong empirical claim for browser-based touring, though it is reported on a single high-end machine without a systematic benchmark across devices.

Usage scenarios

The paper demonstrates two analytical scenarios across text, image, and single-cell data.

Revealing structure. On Fashion MNIST (70K images), the attraction-repulsion tour reveals that a tight cluster of 96 trouser images embedded among shirts and dresses is an artifact of repulsive forces—the points disperse at the ForceAtlas2-like keyframe, and image inspection confirms their silhouettes resemble upper-body garments. Conversely, boundary points persist across the entire spectrum, suggesting boundary placement is more trustworthy than cluster tightness. On 346K CyTOF-profiled immune cells, a spectral Fisher LE tour recovers the known CD4/CD8 division and the CD103/ICOS tissue-resident axis without manual marker specification; CD3 appears only in late frames with low loading, consistent with its constancy across the population. Selecting Tregs and switching to manual mode isolates the ICOS-high, tumor-enriched immunosuppressive subset identified by Mair et al.

Validating embeddings. A little PCA tour over 276K mouse brain cells shows that choroid plexus cells form one cluster in every PCA keyframe but split into two distant UMAP clusters, and blood cells isolated by UMAP as a disconnected island show no comparable separation in the PCA tour—direct evidence that touring PC pairs can distinguish genuine structure from embedding artifacts. On 3M arXiv title-abstract embeddings from four models spanning a 10× range in size (SPECTER2, BGE-M3, Nomic Embed v2, F2LLM-v2-8B), the sequential tour reveals that a prominent compact cluster in F2LLM disperses in all three encoder-based models; analysis of the ~1,200 selected papers shows ~84% are physics education research, indicating F2LLM clusters by discourse style rather than topic, whereas the citation-trained SPECTER2 distributes them among physics subfields. This is the paper's most concrete demonstration that sequential tours surface behavioral differences between models invisible in any single 2D projection.

Limitations and open questions

The paper concedes several constraints. Intermediate frames of sequential embedding tours are geometrically valid interpolations but not interpretable projections of latent structure, so their analytical meaning is limited to the endpoints. The scalability figures are reported only for one hardware configuration, and no user study evaluates whether the fluid mode transitions actually reduce cognitive load compared with fixed-mode tools—the progressive-exploration claim rests on design rationale and usage scenarios rather than controlled evaluation. The tool is agnostic to keyframe provenance, but the paper does not address how analysts should select or prune keyframe sets when the candidate sequence is large. Finally, the validation scenarios rely on qualitative visual judgment of divergences between tours; no quantitative correspondence between tour-observed artifacts and distortion metrics is established.

Conclusion

dtour demonstrates that tour-based exploration of high-dimensional data becomes practical when traversal is fluid and progressive: scrubbing, manual axis manipulation, and grand-tour playback coexist in one interface that scales to millions of points in a browser. Its usage scenarios show that touring across projection directions, hyperparameters, or embedding models can expose artifacts—such as repulsion-induced clusters and model-specific similarity behavior—that no single 2D view reveals, and it leaves open the empirical question of whether such fluid traversal measurably improves analyst performance over conventional fixed-mode tour interfaces.