Comparative Visualisation for High Dimensional Data
Comparative Visualisation for High-Dimensional Data Based on Feature-Wise Binning and Similarity Metrics
My thesis extends AccuStripes for comparative analysis of high dimensional datasets. The system combines interactive distribution views, multiple binning methods and quantitative similarity measures to help compare real and synthetic data.
Overview
This thesis is my most substantial technical project. It extends AccuStripes into a comparative visual analytics system for high dimensional datasets, combining interactive views with quantitative similarity metrics.
The Problem
Comparing high dimensional datasets is hard because differences are not always visible in a single chart. A useful comparison needs to show feature wise distributions, preserve enough structure to support exploration and connect visual patterns with measurable similarity.
What I Built
- Interactive feature wise visualisations for comparing high dimensional datasets.
- Multiple binning strategies for representing distribution structure.
- Similarity based ordering and a similarity network for exploring relationships between datasets.
- Evaluation workflows using real and synthetic datasets.
Binning Methods
The system supports Gaussian kernel based binning, Voronoi binning and Generalised Adaptive Intelligent, or GAI, binning. These methods make it possible to compare how different binning choices affect the visible structure of a dataset.
Similarity Metrics
I integrated Earth Mover's Distance, SSIM and MS SSIM as complementary ways to compare distributions. The thesis uses per dataset and per feature normalisation for visual comparison, while cross dataset per feature normalisation supports metric calculation.
Interface and Interaction
The interface supports feature wise comparison, automatic arrangement by similarity and a similarity network. The goal is to help users move between detailed distribution views and higher level dataset relationships.
Evaluation
The system was evaluated with real and synthetic datasets, including country trends over time, global cross sectional comparisons across years and synthetic distribution comparison. A user study with 19 participants helped examine how people interpreted the visual and metric based results.
Key Findings
- Visual similarity and distribution similarity are related, but they are not always the same.
- Binning methods can change how users perceive structure in the same data.
- Different metrics respond differently depending on the kind of distribution change being compared.
What I Learned
I learned how much design work sits between a mathematical comparison and a useful analytical interface. The project strengthened my understanding of visualisation, metric choice, interaction design and evaluation.