Changed behavior in git vs v2.12

[flying-sheep]

@igordot wrote:

Actually, I have an update on this issue. I tested the new version on a few different data sets. Most of them come out essentially the same, but there is one data set that produced very different results.

This is the result with 2.12:

And this is the with the GitHub version:

The first one is roughly what one would expect and the second one is not. I tried subsampling the number of cells. I am showing 5000 cells above. Reducing it to 4500 yielded a plot that looks normal with both versions, so it seems like the input data is mostly okay.

Do you know what may be causing this?

[flying-sheep]

We have to figure out if this is a bug in my new logic from 1008274 or an effect of the switch to RSpectra in b4e8dad. There were no other relevant changes since then.

1. Please check if it’s a logic bug by installing a version doesn’t have the logic change and test there, i.e. devtools::install_github('theislab/destiny', ref='fbf6f29b4f3eab54095b2c96da27173619babc52')
2. Please check if it’s a random thing by trying several seeds before recreating the diffusion map: set.seed(1), set.seed(2), …

Due to the subsampling making a difference, I suspect that the data just gives unstable results, and using different seeds might make a difference.

[igordot]

I will give it a try. I can also share the input matrix if you'd like.

Regarding subsampling, I was using the same input matrix for both versions.