Source-linked AI summary
t-viSNE: Interactive Assessment and Interpretation of t-SNE Projections
Angelos Chatzimparmpas, Rafael M. Martins, Andreas Kerren
TL;DR
t-SNE projections can be difficult to interpret and may undermine trust because of method-specific pitfalls and hidden algorithmic information. t-viSNE addresses this with coordinated interactive views for exploring projection quality, parameters, neighborhoods, densities, costs, and dimension-pattern relationships; the paper reports promising user-study results and real-data demonstrations.
Problem
t-SNE projections have interpretation and trustworthiness problems, including complex parameter effects, misleading densities and distances, apparent clusters in random data, and difficult-to-assess shapes.
Method
t-viSNE is an interactive tool combining coordinated views to explore t-SNE hyper-parameters, projection accuracy, regional quality, and relationships between projection patterns and original dimensions.
Results
The paper demonstrates t-viSNE with hypothetical scenarios and real-world data, and reports promising results from a user study comparing it with Google’s Embedding Projector.
Takeaways & Limitations
By exposing information normally lost after t-SNE, t-viSNE supports analysts in assessing projections and making their results better understandable.
Takeaways & Limitations
The user study provides initial evidence but may be biased toward t-viSNE and does not represent a large-scale field study with real-world experts in their working environments.
Abstract
from arXiv · showhide
t-Distributed Stochastic Neighbor Embedding (t-SNE) for the visualization of multidimensional data has proven to be a popular approach, with successful applications in a wide range of domains. Despite their usefulness, t-SNE projections can be hard to interpret or even misleading, which hurts the trustworthiness of the results. Understanding the details of t-SNE itself and the reasons behind specific patterns in its output may be a daunting task, especially for non-experts in dimensionality reduction. In this work, we present t-viSNE, an interactive tool for the visual exploration of t-SNE projections that enables analysts to inspect different aspects of their accuracy and meaning, such as the effects of hyper-parameters, distance and neighborhood preservation, densities and costs of specific neighborhoods, and the correlations between dimensions and visual patterns. We propose a coherent, accessible, and well-integrated collection of different views for the visualization of t-SNE projections. The applicability and usability of t-viSNE are demonstrated through hypothetical usage scenarios with real data sets. Finally, we present the results of a user study where the tool's effectiveness was evaluated. By bringing to light information that would normally be lost after running t-SNE, we hope to support analysts in using t-SNE and making its results better understandable.
1 INTRODUCTION
t-SNE can reveal complex high-dimensional patterns, but its complexity creates important trustworthiness and interpretation challenges. t-viSNE addresses these challenges through coordinated interactive views for assessing projection quality and interpreting visible patterns.
- t-SNE creates low-dimensional representations that can capture complex high-dimensional patterns as well-separated clusters.
- t-SNE’s complicated parameter relationships, density distortions, ignored inter-cluster distances, spurious clusters, and difficult-to-assess shapes threaten interpretation and trustworthiness.
- t-viSNE interactively explores t-SNE projections while exposing hidden algorithmic information relevant to the analyzed high-dimensional data.
- Its coordinated views support hyper-parameter selection, projection-accuracy assessment, regional quality investigation, and interpretation through original dimensions.
- The system combines selected views, interaction techniques, and visual mappings in a carefully designed workflow for t-SNE analysis.
- A user study reported promising outcomes, while the tool leaves investigation of misleading clusters from random input for future work.
2 RELATED WORK
Related work addresses dimensionality-reduction exploration, quality assessment, and interpretation through general or specialized tools. t-viSNE differs by targeting t-SNE-specific shortcomings with coordinated views that expose algorithmic information and support quality analysis.
- Dimensionality-reduction tools must support not only scatterplot exploration but also projection-quality assessment and interpretation.
- Single scalar quality scores enable quick overview or automatic selection but cannot capture intricate details of projection quality.
- Interactive interfaces are needed because complex relationships between low- and high-dimensional spaces in nonlinear projections resist simple static mappings.
- Existing tools variously focus on general projection exploration, MDS, specific domains, clustering, or local analysis rather than t-SNE’s particular shortcomings.
- t-viSNE combines coordinated views to expose hidden information from the dimensionality-reduction algorithm and test projection quality.
3 OVERVIEW OF T-SNE
t-SNE transforms high-dimensional neighborhoods into probability distributions and optimizes a low-dimensional distribution to represent them. Its design uses adaptive high-dimensional neighborhood bandwidths, a heavy-tailed low-dimensional distribution, and KL-divergence optimization.
- t-SNE begins with an n × N data matrix and converts pairwise instance distances into probabilities representing high-dimensional neighborhoods.
- The Gaussian bandwidth σ_i varies by instance, allowing high-dimensional neighborhood probabilities to reflect differing local densities.
- The algorithm seeks a low-dimensional probability distribution Q that faithfully represents the joint high-dimensional distribution P, typically in 2-D or 3-D.
- Q uses a one-degree-of-freedom Student’s t-distribution without variable neighborhood density, so differently dense high-dimensional neighborhoods may receive similarly sized low-dimensional areas.
- t-SNE minimizes KL divergence between P and Q using gradient descent for a user-specified number of iterations.
- Computing each point’s cost against all other points has O(N^2) complexity, and the final cost sums the remaining pointwise costs.
4 T-VISNE: A VISUAL INSPECTOR OF T-SNE
t-viSNE is a web-based system designed to address t-SNE’s method-specific interpretability and trustworthiness challenges through coordinated views of projections, data, and hyper-parameters. Its workflow supports projection selection, global and local quality assessment, and investigation of densities, costs, neighborhoods, and dimensions underlying visible patterns.
- t-viSNE targets method-specific shortcomings of t-SNE, including sensitivity to parameters, misleading cluster densities, and difficulty assessing projection quality and shapes.The tool responds to a remaining gap in general-purpose dimensionality-reduction assessment by focusing on t-SNE’s particular problems.
- Goal 1: Hyper-parameter Exploration: A grid-search mode computes 500 projections by varying perplexity, learning rate, and maximum iterations, then presents 25 representative projections with five quality metrics and their QMA.The representatives are selected by clustering projections with K-Medoids using Procrustes distance, while metrics include NH, trustworthiness, continuity, normalized stress, and Shepard diagram correlation.
- Goal 3: Quality: The Shepard Heatmap aggregates pairwise high-dimensional and 2-D distances into cells, mapping cell counts to a single-hue colormap to improve readability for large data sets.It is designed as a less cluttered alternative to a full Shepard Diagram, whose number of pairwise points grows as n^2.
- Goal 3: Quality: Density and Remaining Cost views expose information often lost in t-SNE, helping distinguish clusters and identify regions whose mappings are harder to optimize.In the Iris example, high Remaining Cost occurs where versicolor and virginica points mix; density visualization separates clusters 2 and 3 that t-SNE renders with similar sizes.
- Goals 3–4: Quality and Dimensions: Neighborhood Preservation compares preservation across neighborhood sizes for the full projection and a selected region, while adaptive views help investigate dimensions associated with visible patterns.The Dimension Correlation tool ranks dimensions by absolute correlation with a user-induced ordering and explicitly does not claim that those dimensions cause the observed shape.
5 USE CASES
The use cases show how t-viSNE helps analysts assess projection quality, inspect local distortions, and connect visible patterns to high-dimensional dimensions. Interactive views also support investigating outliers and refining projections when important variation is poorly represented.
- Overall Accuracy: Density inspection revealed that benign breast-cancer cases were more homogeneous in high-dimensional space, while malignant cases were more diverse.The projection’s apparent compactness did not directly reflect the original high-dimensional density structure.
- Investigation of Outliers: Outlier analysis showed that isolated cancer points and mini-clusters could reflect class mixing or groups with identical or highly similar dimension values.This helped distinguish potentially difficult-to-classify cases from projection artifacts.
- Closing the Visual Analysis Loop: Local quality views identified a small high-cost region in diabetes cluster C1 and showed that overlapping C3 points failed to represent internal variation well.The C3 analysis motivated optimizing the selected region rather than accepting the initial projection.
- Interpretation of Clusters: Dimension Correlation linked cluster patterns to specific variables, including weak association with mitoses, strong association with clump thickness, and nearly 70% correlation between the diabetes C2 shape and insulin.Analysts drew polylines along visible patterns and inspected correlated dimensions and their values.
- Closing the Visual Analysis Loop: Re-optimizing C3 with perplexity 10 instead of 50 produced a sparser projection while retaining high quality-metric values and making the cluster easier to explore.The revised projection allowed C3 to be examined without zooming in.
6 USER EVALUATION
The comparative user experiment evaluated t-viSNE against Google’s Embedding Projector through task completion time, task supportiveness, and ICE-T feedback. t-viSNE generally provided stronger task support and significantly higher ICE-T scores, without a significant overall completion-time difference.
- Study design: The study assigned 28 participants to t-viSNE and GEP groups of 14, using six tasks designed around known t-SNE exploration pitfalls.Participants included researchers, students, and practitioners interested in high-dimensional data analysis.
- Completion time: No statistically significant overall completion-time difference was detected between t-viSNE and GEP, although t-viSNE users were faster on Tasks 1 and 5.This addressed RQ1 concerning whether users spent the same time performing tasks in both tools.
- Task supportiveness: t-viSNE received consistently higher tool-supportiveness scores across the proposed tasks than GEP.The authors note that supportiveness reflects subjective user feedback and is one indication of usefulness for exploratory visual analysis.
- Task results: Most t-viSNE participants judged distance preservation as good but imperfect, whereas GEP responses were more scattered toward weaker preservation or uncertainty.The t-viSNE responses aligned with the Shepard Heatmap for the evaluated data set.
- Task results: All t-viSNE participants selected mitoses as the least important dimension affecting cluster shape, while GEP answers were mostly scattered toward uncertainty.The authors describe this as encouraging for consistency, while not claiming that the selected answer is definitively correct.
- ICE-T evaluation: t-viSNE significantly outscored GEP in all four ICE-T components, with its smallest score at 4.75 and GEP receiving many scores below 4.The four components were Insight, Confidence, Essence, and Time; the Time scores were most similar between tools.
7 DISCUSSION
The discussion examines design choices, implementation trade-offs, limitations, and future directions for t-viSNE. It emphasizes adaptive and coordinated views while acknowledging scope, performance, and evaluation boundaries.
- Design Choices: The Shepard Heatmap was designed as a quick, simple overview of projection quality, while the Shepard Diagram exposes more detail and can become confusing on larger data sets.Both approaches remain available so users can choose the scatterplot-based view when desired.
- Design Choices: The tool uses distinct categorical, single-hue sequential, and multi-hue sequential colormaps for labels, the Shepard Heatmap, and the main view.The colormaps were chosen to avoid interference and clarify that the views encode different information.
- Design Choices: t-viSNE offers four Neighborhood Preservation visualizations—bar, difference bar, standard line, and difference line plots—with different trade-offs between comparison, precision, and simplicity.The difference line plot clarifies selection-versus-average differences but sacrifices exact percentage readability.
- Design Choices: Adaptive PCP dynamically rearranges axes after selection, addressing clutter and helping reveal class separation in the 44-dimensional SPECTF data set.It uses PCA as a degree-of-interest function and displays the eight most informative dimensions, unlike the cluttered standard PCP.
- Limitations and Future Work: The implementation prioritized ease of implementation over raw performance by using Barnes-Hut t-SNE instead of newer faster implementations.The authors state that replacing the implementation should be straightforward if necessary.
- Limitations and Future Work: Most t-viSNE views can be reused with other dimensionality-reduction methods, but density analysis is an inherent characteristic of t-SNE.The tool’s algorithm-specific focus could become a limitation if t-SNE were replaced, although many views could be adapted.
- Limitations and Future Work: The user study provides initial evidence of acceptance, consistency, and improvement over another tool, but its tasks may have favored t-viSNE and did not represent a large-scale field study.The authors report positive results while describing the evaluation as promising rather than definitive.
- Limitations and Future Work: A possible future extension would preview point trajectories during another optimization step to help users judge whether t-SNE optimization is complete.This proposal builds on remaining costs as an indicator of projection quality.
8 CONCLUSIONS
t-viSNE is an interactive tool for investigating t-SNE projections by exposing information about projection quality and cluster formation. Its user study found better task support than Google’s Embedding Projector, while completion times were similar.
- Conclusions: t-viSNE exposes t-SNE’s internal information to help users test projection quality and understand cluster formation.It also visualizes densities and areas that are not well-optimized according to t-SNE.
- Conclusions: Participants generally completed the intended analysis tasks with limited training, and feedback indicated better support than Google’s Embedding Projector.The comparison found similar completion times for both tools.