Open ibartomeus opened 6 years ago
Thanks Nacho! Very helpful
The idea of wrapping up an R package to make these analyses easier to the world sounds nice and very fun. Still, I think the paper would gain a lot if we made the analyses ourselves with a lot of networks to examine how much sampling effort is enough, taxonomy, etc. I guess people want to know that beforehand rather than having to make all the work themselves and then discover how much effort would have been enough afterwards...
But of course analysing more networks is a lot more work... The decision is yours. Personally, the idea of an Rpackage combined with analyses of the BEEFUN data sounds alright to me. We're all too busy...
Just a note to consider other measures of goodness of fit apart from R2: http://www.stat.cmu.edu/~cshalizi/mreg/15/lectures/10/lecture-10.pdf#subsection.3.2
So, when are we spending a couple of lovely days back in the field to finish this up? ;-)
We can think on another full day at EBD. The pictures of the conference room make you feel like in cazorla :)
I like the idea of @Pakillo of providing tools to know a priori what sampling resolution is required to properly characterize networks, but in the other hand it requires a lot of effort in compiling published networks, preparing a good data-set, etc. So, because we are all with lots of parallel projects we run the risk that the study falls in the limbo. The idea of @ibartomeus of the framework with an associated R packages looks optimal in this sense and the analyses are considerably advanced. If you will meet soon, let me know. I'll try to be online that day. Hugs from Bariloche, boludos!!
I like the idea of meeting a full day at EBD, My expertise in building an R package is limited, but I can contribute is those tasks I could be useful.
Oscar
On Wed, Sep 20, 2017 at 3:31 PM, Néstor Pérez-Méndez < notifications@github.com> wrote:
I like the idea of @Pakillo https://github.com/pakillo of providing tools to know a priori what sampling resolution is required to properly characterize networks, but in the other hand it requires a lot of effort in compiling published networks, preparing a good data-set, etc. So, because we are all with lots of parallel projects we run the risk that the study falls in the limbo. The idea of @ibartomeus https://github.com/ibartomeus of the framework with an associated R packages looks optimal in this sense and the analyses are considerably advanced. If you will meet soon, let me know. I'll try to be online that day. Hugs from Bariloche, boludos!!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/ibartomeus/dataton/issues/2#issuecomment-330851980, or mute the thread https://github.com/notifications/unsubscribe-auth/AS0oyLg-R7bPD_xixMSWwE5v6WpT4vrnks5skROYgaJpZM4PdqVu .
In general, people liked it, but as a critique, it is a bit system specific. This is my idea to fix that: What we propose is a framework to assess relative sampling intensity and provide the tools to do it with your network. An R package with functions checking smpling completness, effect of taxonomy on sampling, etc...
One figure we need is a plot of R2 between two networks (reference and subsempled) as a function of sampling effort of the subsampled. This is a key function for the package and can use individual subsampling or round subsampling.
Message is "understand your data". I like the idea also of recomending that the question will drive your method. e.g. aggregating species may have a huge impact on metrics like H2 and link evenness, but not on N. Or nestedness as a measure of stability only makes sense when links are as represented as possible. Hence, predicting links may be a good idea.
Add robustness as a metric. Should we add some species level metric??
Finnally: test: