scale selection at margins of domain
Closed this issue · 4 comments
The current scale selection criterion is not reliable at the margins of the domain of Markov scale t due to:
- Cut-off through windowing
- The diagonal count of low values in the NVI(t,t') necessarily leads to smaller values at the margins because the diagonal is smaller. Hence, 1-NVI(t,.) tends to produce anomalous local minima at the margins.
While 1) is perhaps acceptable, Mauricio and me discussed that we should come up with a solution for 2). I will think about this and try a couple of things.
Can you somehow normalize by diagonal size?
Or just don't compute it at the margin, and write in docsctring why it is not reliable, so we don't compute it. User can increase markov time min/max to have larger range of scale detection.
I developed a new algorithm with @mauriciobarahona that seems to work better. The idea is to pool the NVI(t,t') matrix and look at the local minima of its diagonal to identify the block structures. I will create a new branch for my implementation once we merge the Leiden implementation. This new scale selection algorithm will need a slightly modified summary plot, I will try to adapt. Hopefully, I will be quick to update everything for our paper submission and not cause any delays.
Excellent!