Advancing ecological networks: moving beyond binary classification to probabilistic interactions

Abstract

Leveraging trophic interactions to deduce macro-ecological patterns has become a prevalent method, taking advantage of the extensive databases on binary trophic interactions (i.e., prey-predator relationships). However, this binary approach oversimplifies complex ecological dynamics and fails to capture the nuanced structure of food webs. The challenge lies in the scarcity and limited availability of data on non-binary interactions, which are crucial for a more comprehensive understanding of ecological networks. This study explores the use of binary classifiers, particularly the XGBOOST algorithm to address the limitations of traditional binary approaches to prey-predator relationships. By predicting predation probabilities among nine mammalian predators using species traits, my findings demonstrate the classifiers' robust predictive capabilities to binary predictions but also a good correlation between probabilistic predation derived from binary classifiers and observed prey preferences. It also highlighted the importance of selecting informative species traits for predicting interaction, with performance contrastingly superior to null models. Despite a small sample size, this work provides insightful results and sets a foundation for future research to expand these models to broader ecological networks, emphasizing the need for comprehensive prey preference data.