🥪 lunchtime musings

docs/_tex/index.tex

@@ -601,8 +601,8 @@ \section{Network construction is
 network representation in the context of trying to understand the
 feeding dynamics of a seasonal community.
 
-\begin{tcolorbox}[enhanced jigsaw, colframe=quarto-callout-note-color-frame, coltitle=black, titlerule=0mm, breakable, opacityback=0, left=2mm, leftrule=.75mm, toprule=.15mm, rightrule=.15mm, arc=.35mm, title=\textcolor{quarto-callout-note-color}{\faInfo}\hspace{0.5em}{Box 1 - Why we need to aggregate networks at different scales: A
-hypothetical case study}, bottomrule=.15mm, colbacktitle=quarto-callout-note-color!10!white, opacitybacktitle=0.6, bottomtitle=1mm, toptitle=1mm, colback=white]
+\begin{tcolorbox}[enhanced jigsaw, rightrule=.15mm, breakable, colframe=quarto-callout-note-color-frame, opacitybacktitle=0.6, toprule=.15mm, colbacktitle=quarto-callout-note-color!10!white, titlerule=0mm, left=2mm, opacityback=0, bottomtitle=1mm, arc=.35mm, coltitle=black, toptitle=1mm, title=\textcolor{quarto-callout-note-color}{\faInfo}\hspace{0.5em}{Box 1 - Why we need to aggregate networks at different scales: A
+hypothetical case study}, bottomrule=.15mm, leftrule=.75mm, colback=white]
 
 Although it might seem most prudent to be predicting, constructing, and
 defining networks that are the closest representation of reality there
@@ -706,8 +706,9 @@ \subsubsection{Models that predict metawebs (feasible
 testing/benchmarking tools (Poisot, 2023), means that these models can
 be validated and (with relative confidence) be used to construct first
 draft networks for communities for which we have no data (Strydom et
-al., 2022), and are valuable for constructing prehistoric networks
-(Fricke et al., 2022; Yeakel et al., 2014).
+al., 2022), and are valuable for constructing networks where we lack any
+interaction data \emph{e.g.,} prehistoric networks (Fricke et al., 2022;
+Yeakel et al., 2014).
 
 \subsubsection{Models that predict realised networks (realised
 interactions)}\label{models-that-predict-realised-networks-realised-interactions}
@@ -789,10 +790,12 @@ \subsection{Further development of models and
 networks are less complex than they could be, suggesting that there are
 constraints on network assembly. In addition to the more intentional
 development of models we also need to consider the validation of these
-models, There have been developments and discussions for assessing how
+models, there have been developments and discussions for assessing how
 well a model recovers pairwise interactions (Poisot, 2023; Strydom,
-Catchen, et al., 2021) but we lack any clear strategies for benchmarking
-the ability of models to recover structure (Allesina et al., 2008).
+Catchen, et al., 2021), although the rate of false-negatives that may be
+present in the testing data still present a challenge (Catchen et al.,
+2023), and we still lack clear strategies for benchmarking the ability
+of models to recover structure (Allesina et al., 2008).
 
 \subsubsection{At what scale should we be predicting and using
 networks?}\label{at-what-scale-should-we-be-predicting-and-using-networks}
@@ -967,6 +970,11 @@ \section*{References}\label{references}
 \emph{Ecology Letters}, \emph{25}(4), 889--899.
 \url{https://doi.org/10.1111/ele.13966}
 
+\bibitem[\citeproctext]{ref-catchenMissingLinkDiscerning2023}
+Catchen, M. D., Poisot, T., Pollock, L. J., \& Gonzalez, A. (2023).
+\emph{The missing link: Discerning true from false negatives when
+sampling species interaction networks}.
+
 \bibitem[\citeproctext]{ref-cherifEnvironmentRescueCan2024}
 Cherif, M., Brose, U., Hirt, M. R., Ryser, R., Silve, V., Albert, G.,
 Arnott, R., Berti, E., Cirtwill, A., Dyer, A., Gauzens, B., Gupta, A.,

docs/notebooks/model_qualitative-preview.html

@@ -1688,7 +1688,7 @@ <h2 class="anchored" data-anchor-id="visualisation">Visualisation</h2>
 <div id="fig-dendro" class="quarto-figure quarto-figure-center quarto-float anchored">
 <figure class="quarto-float quarto-float-fig figure">
 <div aria-describedby="fig-dendro-caption-0ceaefa1-69ba-4598-a22c-09a6ac19f8ca">
-<a href="../images/dendo.png" class="lightbox" data-glightbox="description: .lightbox-desc-1" data-gallery="quarto-lightbox-gallery-1"><img src="../images/dendo.png" class="img-fluid figure-img"></a>
+<a href="../images/dendo.png" class="lightbox" data-gallery="quarto-lightbox-gallery-1" data-glightbox="description: .lightbox-desc-1"><img src="../images/dendo.png" class="img-fluid figure-img"></a>
 </div>
 <figcaption class="quarto-float-caption-bottom quarto-float-caption quarto-float-fig" id="fig-dendro-caption-0ceaefa1-69ba-4598-a22c-09a6ac19f8ca">
 Dendrogram of the trait table using a hierarchical clustering model, This is based off of the traits table in SuppMat 2)
@@ -2092,7 +2092,7 @@ <h2 class="unnumbered anchored" data-anchor-id="references">References</h2>
     }
   }
 });
-</script>  </div> <!-- /content -->  <script>var lightboxQuarto = GLightbox({"openEffect":"zoom","closeEffect":"zoom","loop":false,"selector":".lightbox","descPosition":"bottom"});
+</script>  </div> <!-- /content -->  <script>var lightboxQuarto = GLightbox({"openEffect":"zoom","closeEffect":"zoom","descPosition":"bottom","selector":".lightbox","loop":false});
 window.onload = () => {
   lightboxQuarto.on('slide_before_load', (data) => {
     const { slideIndex, slideNode, slideConfig, player, trigger } = data;

docs/notebooks/model_quantitative-preview.html

@@ -952,7 +952,7 @@ <h2 class="unnumbered anchored" data-anchor-id="references">References</h2>
     }
   }
 });
-</script>  </div> <!-- /content -->  <script>var lightboxQuarto = GLightbox({"openEffect":"zoom","descPosition":"bottom","loop":false,"selector":".lightbox","closeEffect":"zoom"});
+</script>  </div> <!-- /content -->  <script>var lightboxQuarto = GLightbox({"loop":false,"closeEffect":"zoom","descPosition":"bottom","openEffect":"zoom","selector":".lightbox"});
 window.onload = () => {
   lightboxQuarto.on('slide_before_load', (data) => {
     const { slideIndex, slideNode, slideConfig, player, trigger } = data;

index.qmd

@@ -154,7 +154,7 @@ There is a bit of a 'point of conflict' between those calling for 'pixel perfect
 
 ### Models that predict metawebs (feasible interactions)
 
-This is perhaps the most developed group of models; with a variety of approaches having been developed that typically determine the feasibility of an interaction based on the trait compatibility between predator and prey (*i.e.* their evolutionary compatibility) to determine 'feeding rules' [@morales-castillaInferringBioticInteractions2015]. These feeding rules are broadly elucidated in two different ways; mechanistic feeding rules can be explicitly defined and applied to a community [*e.g.,* @shawFrameworkReconstructingAncient2024; @dunneCompilationNetworkAnalyses2008] or they are inferred from a community for which there is interaction data and the 'rules' are then applied to a different community [*e.g.,* @strydomGraphEmbeddingTransfer2023; @pichlerMachineLearningAlgorithms2020; @strydomFoodWebReconstruction2022; @caronAddressingEltonianShortfall2022; @llewelynPredictingPredatorPrey2023; @desjardins-proulxEcologicalInteractionsNetflix2017; @eklofSecondaryExtinctionsFood2013; @cirtwillQuantitativeFrameworkInvestigating2019]. The fundamental difference between these two model groups is that 'mechanistic models' rely on expert knowledge and make assumptions on trait-feeding relationships, whereas the 'pattern finding' models are dependent on existing datasets from which to elucidate feeding rules. These models are useful for determining all feasible interactions for a specific community, and owing to the availability of datasets [*e.g.,* @poelenGlobalBioticInteractions2014; @poisotMangalMakingEcological2016; @grayJoiningDotsAutomated2015], as well as the development of model testing/benchmarking tools [@poisotGuidelinesPredictionSpecies2023], means that these models can be validated and (with relative confidence) be used to construct first draft networks for communities for which we have no data [@strydomFoodWebReconstruction2022], and are valuable for constructing prehistoric networks [@yeakelCollapseEcologicalNetwork2014; @frickeCollapseTerrestrialMammal2022]. 
+This is perhaps the most developed group of models; with a variety of approaches having been developed that typically determine the feasibility of an interaction based on the trait compatibility between predator and prey (*i.e.* their evolutionary compatibility) to determine 'feeding rules' [@morales-castillaInferringBioticInteractions2015]. These feeding rules are broadly elucidated in two different ways; mechanistic feeding rules can be explicitly defined and applied to a community [*e.g.,* @shawFrameworkReconstructingAncient2024; @dunneCompilationNetworkAnalyses2008] or they are inferred from a community for which there is interaction data and the 'rules' are then applied to a different community [*e.g.,* @strydomGraphEmbeddingTransfer2023; @pichlerMachineLearningAlgorithms2020; @strydomFoodWebReconstruction2022; @caronAddressingEltonianShortfall2022; @llewelynPredictingPredatorPrey2023; @desjardins-proulxEcologicalInteractionsNetflix2017; @eklofSecondaryExtinctionsFood2013; @cirtwillQuantitativeFrameworkInvestigating2019]. The fundamental difference between these two model groups is that 'mechanistic models' rely on expert knowledge and make assumptions on trait-feeding relationships, whereas the 'pattern finding' models are dependent on existing datasets from which to elucidate feeding rules. These models are useful for determining all feasible interactions for a specific community, and owing to the availability of datasets [*e.g.,* @poelenGlobalBioticInteractions2014; @poisotMangalMakingEcological2016; @grayJoiningDotsAutomated2015], as well as the development of model testing/benchmarking tools [@poisotGuidelinesPredictionSpecies2023], means that these models can be validated and (with relative confidence) be used to construct first draft networks for communities for which we have no data [@strydomFoodWebReconstruction2022], and are valuable for constructing networks where we lack any interaction data *e.g.,* prehistoric networks [@yeakelCollapseEcologicalNetwork2014; @frickeCollapseTerrestrialMammal2022]. 
 
 ### Models that predict realised networks (realised interactions)
 
@@ -168,7 +168,7 @@ Although we identify mechanisms that determine species interactions in @sec-proc
 
 ## Further development of models and tools
 
-There has been a suite of models that have been developed to predict trophic links, however we are lacking in tools that are explicitly taking into consideration estimating both the feasibility as well as realisation of links, *i.e.,* both interactions and structure simultaneously [@strydomRoadmapPredictingSpecies2021]. This could be addressed either through the development of tools that do both (predict both interactions and structure), or to develop an ensemble modelling approach [@beckerOptimisingPredictiveModels2022]. Alternatively the development of tools that will allow for the downsampling of metawebs into realised networks [*e.g.,* @roopnarineExtinctionCascadesCatastrophe2006], although deciding exactly what is driving differences between local networks and the regional metaweb might not be that simple [@saraviaEcologicalNetworkAssembly2022]. Probably also something that aligns with trying to predict interaction strength - because that would be the gold standard. Probably also worth just plainly stating that feasibility of developing a model that is both broadly generalisable, but also has local specificity is probably not attainable [@stoufferAllEcologicalModels2019], and more specifically the potential use in models untangling/identifying the different processes that shape interaction networks [@songRigorousValidationEcological2024], *e.g.,* @curtsdotterEcosystemFunctionPredator2019 showcasing the use of models to disentangle the drivers of community function and @strydomSVDEntropyReveals2021 who identified that networks are less complex than they could be, suggesting that there are constraints on network assembly. In addition to the more intentional development of models we also need to consider the validation of these models, There have been developments and discussions for assessing how well a model recovers pairwise interactions [@strydomRoadmapPredictingSpecies2021; @poisotGuidelinesPredictionSpecies2023] but we lack any clear strategies for benchmarking the ability of models to recover structure [@allesinaGeneralModelFood2008].
+There has been a suite of models that have been developed to predict trophic links, however we are lacking in tools that are explicitly taking into consideration estimating both the feasibility as well as realisation of links, *i.e.,* both interactions and structure simultaneously [@strydomRoadmapPredictingSpecies2021]. This could be addressed either through the development of tools that do both (predict both interactions and structure), or to develop an ensemble modelling approach [@beckerOptimisingPredictiveModels2022]. Alternatively the development of tools that will allow for the downsampling of metawebs into realised networks [*e.g.,* @roopnarineExtinctionCascadesCatastrophe2006], although deciding exactly what is driving differences between local networks and the regional metaweb might not be that simple [@saraviaEcologicalNetworkAssembly2022]. Probably also something that aligns with trying to predict interaction strength - because that would be the gold standard. Probably also worth just plainly stating that feasibility of developing a model that is both broadly generalisable, but also has local specificity is probably not attainable [@stoufferAllEcologicalModels2019], and more specifically the potential use in models untangling/identifying the different processes that shape interaction networks [@songRigorousValidationEcological2024], *e.g.,* @curtsdotterEcosystemFunctionPredator2019 showcasing the use of models to disentangle the drivers of community function and @strydomSVDEntropyReveals2021 who identified that networks are less complex than they could be, suggesting that there are constraints on network assembly. In addition to the more intentional development of models we also need to consider the validation of these models, there have been developments and discussions for assessing how well a model recovers pairwise interactions [@strydomRoadmapPredictingSpecies2021; @poisotGuidelinesPredictionSpecies2023], although the rate of false-negatives that may be present in the testing data still present a challenge [@catchenMissingLinkDiscerning2023], and we still lack clear strategies for benchmarking the ability of models to recover structure [@allesinaGeneralModelFood2008].
 
 ### At what scale should we be predicting and using networks?