Hello group
I was running Songbird to identify differently abundant features on my data. When checking how Songbird model fits my dataset, i think my graphs have a nice shape (although Loss graph could me improved).
!qiime songbird multinomial
!qiime songbird summarize-single
!qiime tools view Run1/Songbird/regression-summary.qzv
When running the null model i obtain a pseudo Q-squared of 0.421411, but the orange line on the graphs doesn't look right to me.
!qiime songbird multinomial
and the null model's
!qiime songbird summarize-paired
!qiime tools view Run1/Songbird/paired-summary.qzv
Any ideas why this could be happening? I know the Q-squared should be positive, but is it okay with a graph like that?
mortonjt
May 19, 2023, 2:45pm
2
Did you include the wrong plots? I don't see the orange trend.
Sorry, attached the wrong graph.
mortonjt
May 22, 2023, 10:43am
4
hmm. I have not seen a plot like that before -- looks like something is off with the summary-interval (i.e. the timing of recording the metrics is off ...)
But given the the cross-validation score has flat lined for both models, I'd say the run is solid, so you should be able to run with it.
Hi i realized i made a mistake on my script for the null model (--p-training-column Testing) , as i was not using the Testing column for the first set of graphs.
Now i am wondering which aspect should i prioritize, graph shape or Q-squared value
!qiime songbird multinomial
!qiime songbird summarize-single
!qiime songbird multinomial
and the null model's
!qiime songbird summarize-paired
When setting --p-learning-rate to 0.0001, loss function looks a bit better but Q-squared reduces to 0.023946.
mortonjt
May 23, 2023, 1:50pm
7
I'd prioritize Q2. So I think your earlier model is better.
