VC: Star Coordinates, RadViz, GrandTour
In this post, we will talk about visualization techniques that are not a scatterplot or parallel coordinates. The visualization techniques we will discuss here are star coordinates, RadViz, and GrandTour.
Star Coordinates
The attributes of the data will span in the 2D space and calculate the vector sum each other and the sum of the vector will be the embedding data point representing the original data. It looks great and reasonable but there is a significant drawback the embedding data point can be reached other data points also. Therefore, we need to build interaction with the graph.
The prior information about our original data can tell us some attributes are not important for us or clusters desired to observe. We can interact with the attribute vectors to achieve this prior information. Rotation helps to find the desired clusters and scaling can remove the significance of some attributes.
RadViz
This is an alternative way to embed the p-dimensional data in two dimensions. Each attribute represented by anchor points is often uniformly distributed on a circle, e.g. above picture. di,j is the spring force to calculate equilibrium between anchor points and original data points and it is defined by each min-max normalized attributes.
The interaction fo RadViz is re-ordering anchor points because the order of anchor points can make different images.
X-RadViz is suggested to increase interpretability and reduce overlap of points. Differences are using spreading out anchor points along with line segments and adding partial spring lines indicating its own point.
Grand Tours
Grand tours make the animation of each 2D projections of a multi-dimensional space. There are 4 rules for this animation:
- Continuous to allow for visual tracking (Move continuously)
- Dense in the space of all projections (Come with similar one)
- Becoming dense rapidly (With no delay)
- Uniform. (Projection should uniform)
This post is published on 9/3/2020
