Through the Random Forest: Prediction and Explanation of Stage Structured Trophic Dynamics

Background/Question/Methods

As modeling tools and approaches become more advanced, increasingly complex ecological models require more involved methods of analysis. One powerful tool for exploring this new complexity is the application of machine learning (ML) approaches to ecology. While these ML approaches are powerful, their results may suffer from well-known trade-offs between predictive and explanatory power. One common source of increasing model complexity is more explicit modeling of complex organismal ontogeny. The history of incorporating ontogeny through stage-structured models into population ecology has made clear the importance of demographic heterogeneity in influencing ecological dynamics. This has been particularly well-implemented in the study of plant populations. Building off the history of structured models in plant ecology, we integrate trophic interactions across both single- and multi-stage herbivores to investigate how plant ontogeny mediates trophic dynamics at the most basal level of consumer-resource interactions, herbivory. We employ a small, but empirically rooted, ontogenetic modeling framework to investigate how ontogeny functions in trophic dynamics and how the complexity that arises can be captured by both ML models and more simple statistical models grounded in ecological principles.

Results/Conclusions

Our analysis finds categorical differences in the effect of ontogeny on basic consumer-resource dynamics stemming from interacting plant demographics and trophic dynamics. These categorical differences result from high context-dependence in density-dependent maturation and consumption rates, which produces a wide range of dynamic outcomes. In order to untangle this context-dependence, a random forest (RF) model was trained on stability results and produced highly accurate predictions of trophic dynamics. Our machine learning process shows a high degree of interactivity emerging from ontogenetic considerations; specifically, we found that the effect of any single parameter on model behavior could change substantially based on the values of our other parameters. The high context-dependence detected by our RF model was not able to elucidate the key mechanisms driving ecological dynamics. In an effort to achieve explanatory power beyond the results of our RF model, we were able to reduce the complexity of model results to 3 ecologically founded, mechanistic variables (density of seeds germination, density of seedling maturation, and the ratio of consumed to non-consumed plant density). These variables inform a simple linear model whose results rival the predictability achieved by our ML approach while also explaining the ecological mechanism driving observed dynamics.