Grandscatter

To deal with a 14-dimensional space, visualize a 3-D space and say 'fourteen' to yourself very loudly. Everyone does it. - Geoffrey Hinton

Grandscatter is a scatterplot library for interactively exploring high-dimensional datasets as simple projections in two- or three-dimensional space.

Grandscatter is written in TypeScript. This project also provides a Python-based widget for Jupyter.

In grandscatter, the projection plane (or space) is stationary and you travel within data space (i.e., change the visible projection) by manipulating the data axes around the origin. You can rotate and reflect the axes manually, or orchestrate high-dimensional "tours" programmatically.

Grandscatter supports two projection modes:

• orthographic: a flat projection where distances within the 2D canvas are the same as in the projected 2D subspace
• perspective: a 3D projection view, including foreshortening of distal points to improve depth perception

Current supported interactions:

• Drag an axis handle to rotate it
• Option/alt+click an axis handle to flip its orientation
• Shift+drag to lasso-select points

This library was originally based on the Grand Tour implementation in Li et al., 2020. Visualizing Neural Networks with the Grand Tour. Our original implementation was eigen-tour.

Development

To build grandscatter:

pnpm install
pnpm build

To serve the example pages locally:

pnpm dev

For widget development, create a Python dev envrionment:

cd python
uv sync

Run the dev server to watch and rebuild the widget assets:

pnpm dev:widget

Open JupyterLab or other notebook environment:

ANYWIDGET_HMR=1 uv run juptyer lab

Or execute %env ANYWIDGET_HMR=1 in a cell to activate anywidget's hot module replacement.