Closed snotskie closed 7 months ago
Slight tweak, staged in https://github.com/snotskie/EpistemicNetworkAnalysis.jl/commit/063f9dcff2012bcebf82c7dcb57c8d6be5e0f5d6
Using the mean vector (normalized) instead of 1. Here we see a perfectly straight line, not because this makes things straighter, but because the method I was using to "show" the "umbrella" looked for the mean line, not 1
The impact on SVD was minimal in this case, but the choice of the mean vector instead of 1 might be more justifiable. The mean vector tends to be important for downstream steps, and there is no guarantee that 1 will be near any data
The impact on non-SVD rotations is shown below
Before
After
Add an alternative to sphere normalization, which does not create an "umbrella" axis
One alternative is "line normalization". This normalizes units to the plane perp to 1, adjusting distances from 1 such that the new linear distance is equal to the original arc distance
As a proof of concept:
Before:
After:
Above, I've forced the x-axis to align moreorless with the "umbrella" axis. Notice the drop in variance explained as well as the change in shape
Here I've run a stock SVD rotation, before on the left, after on the right. Notice the change in the shape. The "bend" on the left was due to the normalization technique, not a meaningful feature of the data. The image on the right removes the bend, showing more meaningful variance