Projection Ensemble: Visualizing the Robust Structures of Multidimensional Projections



Myeongwon Jung, Jiwon Choi, Jaemin Jo

 Room: 105

2023-10-25T00:45:00ZGMT-0600Change your timezone on the schedule page
2023-10-25T00:45:00Z

Exemplar figure, described by caption below — Projection Ensemble recognizes robust structures in multidimensional projections. A) A randomly initialized t-SNE projection for the MNIST dataset. The viewer may interpret groups of points (a) and (b) as individual clusters, which, in fact, have intricate structures. B) Projection Ensemble visualizes two robust structures identified by extracting common subgraphs between ten randomly initialized projections. This reveals that (a) actually consists of two entangled structures (group (c) and the other points), and a subgroup of the cluster (b) is found to be closer to a distant cluster (see (d)). C) The ground-truth class labels are shown as the color of points.

Fast forward

Full Video

Keywords

Human-centered computing—Visualization—Visu- alization systems and tools

Abstract

We introduce Projection Ensemble, a novel approach for identifying and visualizing robust structures across multidimensional projec- tions. Although multidimensional projections, such as t-Stochastic Neighbor Embedding (t-SNE), have gained popularity, their stochastic nature often leads the user to interpret the structures that arise by chance and make erroneous findings. To overcome this limitation, we present a frequent subgraph mining algorithm and a visualization interface to extract and visualize the consistent structures across multiple projections. We demonstrate that our system not only identifies trustworthy structures but also detects accidental clustering or separation of data points.