Predicting missing links in food webs using stacked models and species traits

Area of Science:

• Ecology
• Computational Biology
• Network Science

Background:

• Ecological networks, like food webs, are vital for understanding ecosystem complexity.
• Existing ecological networks often lack complete species interaction data.
• Inferring missing links computationally aids ecological fieldwork and process investigation.

Purpose of the Study:

• To develop and evaluate a stacked generalization approach for predicting missing links in ecological food webs.
• To assess the method's ability to integrate structural and trait-based ecological predictions.
• To analyze the performance and drivers of link predictability across diverse ecosystems.

Main Methods:

• Employed a stacked generalization machine learning technique to combine diverse ecological predictors.
• Accounted for ecological assumptions, including the directionality of species interactions.
• Validated the method on synthetic food webs and a global database of 290 real-world food webs.

Main Results:

• The stacked generalization method demonstrated high accuracy, often achieving near-perfect performance on real food webs.
• Performance improved when utilizing both species traits and network structure information.
• Predictability of ecological links was found to be influenced by ecosystem type and network characteristics.
• The model's predictions were primarily driven by a select group of ecologically meaningful predictors.

Conclusions:

• Stacked generalization offers a robust and broadly applicable computational tool for predicting missing ecological interactions.
• This approach enhances our ability to study and understand complex ecological systems.
• The findings highlight the importance of integrating multiple data types (traits and structure) for accurate ecological network inference.