uber / manifold

A model-agnostic visual debugging tool for machine learning
Apache License 2.0
1.65k stars 117 forks source link

Visualization improvement for heatmaps in geo-spatial features #36

Open Firenze11 opened 5 years ago

Firenze11 commented 5 years ago

Currently we have 2 ways to show numerical values over spatially distributed data:

  1. Overlapping heatmaps, used for displaying two data slices on a map:

    geo feature view lat-lng
  2. Hexagon binned heatmaps, used for displaying spatially aggregated 1-Dimensional data (e.g. values of one feature):

    geo feature view hex id

The drawbacks are:

  1. Overlapping heapmaps might occlude information
  2. It's hard to show multiple dimensions at the same time

We need to invent new visual encodings to fix it.

Firenze11 commented 4 years ago

Inspirations: https://pair-code.github.io/interpretability/uncertainty-over-space/