Master's thesis · Sep 2025 - Mar 2026

Comparative Visualisation for High-Dimensional Data

An interactive system for comparing feature distributions across several datasets.

The system helps users compare high-dimensional datasets feature by feature, then inspect broader relationships through metric-based ordering and a similarity network.

ReactTypeScriptD3.jsPythonGaussian binningVoronoi binningGAISSIM

Methods

Gaussian, Voronoi and GAI binning

Metrics

SSIM, MS-SSIM and Earth Mover's Distance

Evaluation

Three use cases and 19 participants

Inside the project

Real screenshots and exported results from the working system.

AccuStripes matrix comparing country datasets across child mortality, fertility, GDP per capita and life expectancy. — The comparison view shows feature distributions for United States, India, Germany and Brazil using the same binning configuration.

AccuStripes matrix comparing global cross-sectional datasets from 2020, 2000 and 1980. — Global year datasets are arranged side by side to compare how feature distributions change across 2020, 2000 and 1980.

AccuStripes matrix comparing four synthetic datasets across five feature distributions. — Synthetic datasets make it easier to test whether binning choices reveal or hide known distribution differences.

Similarity network connecting country datasets with SSIM, MS-SSIM, VAICo and Earth Mover's Distance values. — The similarity network connects datasets and exposes metric-specific relationships through edge colour and width.

Bar chart of aggregated user study responses with mean scores between 3.90 and 4.25 on a five-point Likert scale. — Aggregated Likert-scale responses from the user study show positive ratings for interpretation, trend identification, clarity, understanding and trust in the binning techniques.

The problem

Comparing several high-dimensional datasets is difficult because the differences may appear in shape, density, peaks or gaps within individual features. Separate charts also force users to perform too much mental alignment.

What I built

Aligned feature-wise AccuStripes matrices for side-by-side comparison.
Gaussian kernel based, Voronoi and GAI adaptive binning.
SSIM, MS-SSIM and Earth Mover's Distance for complementary similarity views.
Automatic arrangement by similarity and an interactive similarity network.

Normalisation

For display, each dataset and feature uses its own normalised range so the visible structure is readable. For metric calculation, each feature is normalised across datasets so the comparison takes place in a shared domain.

Evaluation

I tested the system with country trends over time, global cross-sectional snapshots across years and synthetic distributions with known differences. A user study with 19 participants assessed interpretation, comparison, sensitivity and usefulness.

What the results showed

Two datasets can look similar while their distributions differ.
The selected binning method changes which structure becomes visible.
SSIM, MS-SSIM and Earth Mover's Distance respond to different kinds of similarity.
A single score cannot explain every relationship between datasets.

What I learned

The strongest part of the project was not one algorithm or chart. It was connecting data processing, visual design, interaction and evaluation into one system that people could actually test.