Species Interaction Network
To construct separate Fall and Spring nondirectional species interaction networks, we adopted the methods of EMtree as proposed by Momal et al. (2020). Generally, EMtree uses both (i) PLN models to represent the joint distribution of species abundances and (ii) spanning trees to create species interaction networks, where a spanning tree is defined as a subset of a network that connects all nodes with the minimum number of possible connections (Dale & Fortin 2014).
All network inference approaches try to reconstruct the underlying true network configuration, but are impeded by the vast number of configurations that a network may have (Momal et al. 2020). To overcome this issue, EMtree employs a tree-based approach to reduce the number of possible network configurations. EMtree constructs fitted species interaction networks by averaging across the spanning trees and employs an advanced tree-based algorithm to maximize the likelihood of the inferred species interactions from the PLN models. The EMtree approach combines both pairwise potential direct (e.g. predator-prey interactions) and indirect (e.g. indirect competition) interactions, represented as a single undirected connection between species’ nodes. Each connection within the network was weighted with a value between zero and one, representing the conditional probability of each connection being part of the true underlying network. We assumed that if an interaction had a non-zero weight, it existed in the network.
To create the network, we had to select a minimum threshold as a cut-off for inferring species connections. This threshold can be used as a metric for assessing the reliability of connections with higher thresholds indicating higher reliability. We evaluated our network using consecutive thresholds of 0.2 between 0.1 and 0.9 (0 is the minimum possible threshold assuming virtually all connections and 1 is the maximum possible threshold producing 0 connections). We increased network robustness by iteratively resampling the network 100 times. The EMtree approach was implemented using the EMtree package (Momal et al. 2020) in R version 3.6.0 (R Core Team 2020).