+1

Maybe add a reference or explain this is true. Isn't it contrary to assuming 1s and 0s are all true when finding the cutoff in Fig.2?

I guess I do not understand what you did for this step and why. I am confused by the link between the trait space confidence intervals and this new threshold. I am also not sure to understand how the threshold was used. Note that I haven't look at Cirtwill and Hamback yet

So after applying the threshold, an interaction with a probability > 0.08 is transformed into an interaction? Can we still talk about "Probability of an interaction" in the probabilistic metaweb P ?

How did you calculated success rate ? If I understand correctly you have probabilities vs ones. Maybe you need to explain what a success mean in that case. Is it probabilities > a certain threshold = success? Or is it successes across all 2*10^5 network? If so, I am not sure to understand what "33 of which were not predicted by the model".

I am not sure to understand the figure. (Left) Red is the proportion of non-zero for each cutoff, but I am not sure to understand where the blue comes from. Also you y label "Proportion of links left" does not apply to the right panel.

Could we use the 0s in Strong & Leroux as it is suppose to document all interactions? If not, maybe a sentence to explain why as you did for Globi.

Same as above

Is it the success rate prior to or after re-using the knowledge from Europe?

fig.3

(fig.3)?

accuracy as in the proportions of correct predictions? or is it the Youden's J? I think, it would be interesting to have the J statistic as well since it is used to determine the cutoff. I agree with Ben that a sentence to describe Youden's J and why you used it would be helpful.

I find it a bit confusing how the rest of the paragraph explain how representational approaches are more easily transferable than other approaches

Maybe a few words on why Canada?

Took me a few read to understand the structure of the sentence. to infer? Maybe use a more precise term (metaweb?) instead of "this representation".

Maybe "presented"? as the paper is not a "review" of the different methods, but an overview of how this can be done.

Make it as 7: Department of BIology, McGill University , Montréal, Canada

Q:Why should we predict interactions networks? A: Because we should!

As a reader, I'd be a bit disappointed by this answer. Maybe something like : Because we almost can, and because we definitely need to.

I think you mean: A neutral approach to probabilistic interactions would rely only on the effect of abundances and assume trait-matching would have no effect?

interaction

C is already used in the ms for the "cooccurrence matrix" in the proof-of-concept

I'd continue with the same structure, something like:

Priors on parameters P(\tetha), which describe the modelers belief about the value of the parameters.

I think you can delete this. We highlight the areas in another section of the ms.

missing ()

If we are talking about coexistance and not interaction, Dunne 2006 is not really relevant

I find it confusing how co-occurrence and interactions are mixed.The first half of the sentence seems to indicate that C is built with interactions, while the second half seems to indicate that it is built with co-occurrence.

If C is built with interactions, maybe we should name it adjacency matrix A ?

This abbreviation has not been introduced. Replace by machine learning

nor

you say twice there have been calls. Maybe simplify the sentence with something like: through probabilistic species pool, and more importantly ...

missing space

missing words

Semi- by curiosity, semi- because it might be better to explicitly say it, but how was accuracy calculated? It is not clear if correctly predicted absences are used? As stated before, there are no true negative.

Maybe I'm wrong, but this applies only to mathematical models?

The sentence is very hard to read. too many long noun strings.