VisCoDeR
A tool for Visually Comparing Dimensionality Reduction Algorithms
Abstract
We propose VisCoDeR, a tool that leverages comparative visualization to support learning and analyzing different dimensionality reduction (DR) methods. VisCoDeR fosters two modes. The Discover mode allows to qualitatively compare several DR results by juxtaposing and linking the resulting scatterplots. The Explore mode allows for analyzing hundreds of differently parameterized DR results in a quantitative way. We present case studies that show that our approach helps to understand similarities and differences between DR algorithms.
Video
High-res images
Screenshots of the Discover mode — (a): linked textual description; (b): juxtaposed DR results, (i) component plane in the background, (ii) proximity view in the background, (iii) ray crossed by an other ray; (c): distribution of all data (grey), selected point and data of its class (pink) along input dimensions.
Screenshot of the Explore mode — (a): color legends of DR algorithms and their parameterizations; (b): control over the meta t-SNE; (c): direct encoded meta view with 1004 DR results; (d): clicking a dot in the meta view displays the DR result in detail.
Screenshot of the juxtaposition of the DR results of the MNIST dataset — For the MNIST dataset, t-SNE offers the most correct result regarding labels as ground truth. The different digits are clearly separated based on their image pattern, while all DR ignore their class labels. Outlying digits can also be explained by referring to their snippet images showing the original digit data.
Screenshot of three types of overview — (left): parallel coordinates; (middle): scattermatrix combined with density plots; (right): direct 3d visualization.
Downloads
Datasets
Spotify Iris MNIST S-shape Swissroll Waves
DR algorithms
Tool
Link to VisCoDeRAuthors
René Cuturarene.cutura@gmail.com
Stefan Holzerst.holzer@gmx.net
Michaël Aupetitmaupetit@hbku.edu.qa
Michael Sedlmairmichael.sedlmair@univie.ac.at