search

biocore / mmvec

Neural networks for microbe-metabolite interaction analysis
BSD 3-Clause "New" or "Revised" License
118 stars 51 forks source link

Obvious instability warnings #85

Closed wasade closed 5 years ago

wasade commented 5 years ago

The produced biplots can sometimes have surprising proportions explained, such that axis 1 < axis 2 < axis 3 or where the proportion explained for an axis is > 100%. It would be nice if mmvec generated a warning when these scenarios are observed in the results and to suggest that the user may not want to use the output.

mortonjt commented 5 years ago

@wasade - do you have an explicit ordination results file that shows these wonky results?

I'm not able to reproduce these results - it'll be great if we could patch it before the next release.

wasade commented 5 years ago

Franck does -- unsure what his GH handle is, I'll follow up on the relevant email thread and ask him to provide context here.

mortonjt commented 5 years ago

ok, got the ordination file from Franck. Below are the eigenvalues / proportion explained

Eigvals 3
0.3058593288174586  2.3895480775642652  9.001766424679664

Proportion explained    3
0.0010773261702244183   0.06575598584302014 0.9331666879867555

These values looks reasonable to me, but the emperor plot doesn't reflect this at all. Below is the file. Not sure why this is the case - I wonder if it has to do with the ordering of the eigenvalues in #86 . Let's try to knock out that issue first and see if this error remains.

CC @ElDeveloper just in case.

File is below latent_dim_3_input_prior_1.00_output_prior_1.00_beta1_0.85_beta2_0.90_ordination.txt

mortonjt commented 5 years ago

Thanks @wasade @FranckLejzerowicz this should be addressed now.