Predicting interaction frequency in plant-pollinator networks
Preprint
Abstract
Flowers and their pollinators represent a bipartite interaction system, whose links are hypothesised to be related to species traits. To explore whether we can predict the weight of this link, i.e. the frequency of interactions, in an validation network, we analysed 14 studies of pollinator-flower visitation network from around the world. We used information on species abundances, their traits and their phylogenetic (for plants) or taxonomic (for animals) position as predictors of interaction frequency, and fitted different statistical modelling approaches. We expected to see prediction quality on validation data to decay with spatial and temporal distance to the training networks. Similarly, we expect that changes in pollinator or plant composition will negatively affect predictive performance. Using the best-predicting modelling approach (randomForest), we indeed see a slight decay in predictive quality with plant and pollinator compositional distance. Temporal distance played little role, although predictions for one year ahead (or back) were better than across the season or across multiple years. The overall predictive power of our models was low (Spearman's ρ ≈ 0.4), suggesting a very noisy system. Also, the most important predictor was abundance, as revealed by a parameter-free benchmark model that only used the cross-product of abundances to predict interaction frequency. Trait and phylogenetic information did not substantially improve predictive performance beyond abundance-based predictions. Across all studies, we failed to confirm a substantial contribution of ecological characteristics to pollinator-flower interaction frequency. One reason why predictions were relatively poor is that sampling effort is not standardised, and thus networks differed substantially in the observed number of interactions, network size, and interaction density. Also the pooling of networks across space or across time may have diluted preferences in the data, reducing their explanatory value. Finally, the majority of species in each network are rare, and the interaction information they provide may be much less relevant that that of common species. At present, we conclude that the frequencies of interactions are very difficult to predict, and using traits we cannot really do better than simply using abundance information.
What the paper shows and why it matters
(AI-generated)
Network ecologists assume that species traits causally shape who visits which flower: interactions are more likely when traits match. The idea is reasonable: some flowers are clearly adapted to animal pollination. If it holds, a model that learns from traits and abundances should predict interaction frequency, that is how often a pollinator visits a plant, in a network at a new site or time. Earlier work mostly predicted whether a link exists (Terry and Lewis 2020); the paper knows of only two earlier attempts at frequency, both technical tests (Pichler et al. 2020; Benadi et al. 2022). The problem it takes on is whether traits, and phylogeny as a stand-in for traits that are hard to measure, improve predictions in new networks beyond species abundance alone.
The paper finds that interaction frequency is hard to predict, and that abundance does practically all of the predicting while traits and phylogeny add little. It tested this on 14 published studies of plant-pollinator networks recorded repeatedly in space or time, predicting held-out networks. Random forest, gradient boosting and a trait-matching model all reached only modest skill; for random forest, the best, the rank correlation (Spearman’s ρ; 1 would be perfect) was 0.42 in time and 0.27 in space. In space all three matched a simple benchmark with nothing to fit, the product of plant and pollinator abundances; in time, random forest and gradient boosting exceeded it only slightly. For pollinators, the abundances were mostly the network’s own visit totals, which the paper acknowledges is circular. Predicting a whole left-out study failed: a model trained on the other 13, using phylogeny only, had virtually no skill (mean ρ of 0.10, not distinguishable from zero).
The paper suggests that, for now, interaction frequency is best treated as driven mainly by abundance, with little generalisable trait signal and much chance variation. It does not rule out trait matching, which may be hard to detect in visitation data, since a visit does not necessarily result in pollination. Scent and vision traits, rather than larger data sets or more elaborate methods, may, the paper suggests, improve predictions; it names unstandardised sampling, pooling of networks and the rarity of most species as possible limits. This is a preprint and has not been peer reviewed.
References
- Terry & Lewis (2020). Finding missing links in interaction networks. Ecology 101, e03047. doi:10.1002/ecy.3047
- Pichler et al. (2020). Machine learning algorithms to infer trait‐matching and predict species interactions in ecological networks. Methods in Ecology and Evolution 11, 281–293. doi:10.1111/2041-210x.13329
- Benadi et al. (2022). Quantitative Prediction of Interactions in Bipartite Networks Based on Traits, Abundances, and Phylogeny. The American Naturalist 199, 841–854. doi:10.1086/714420