Decomposing predictability to identify dominant causal drivers in complex ecosystems.

Area of Science:

• Ecology
• Complex Systems Science
• Data Science

Background:

• Ecosystem dynamics involve complex physical, biological, and chemical processes.
• Stochasticity and nonlinearity in ecosystem data challenge traditional causal inference methods.
• Bridging dynamical and statistical approaches is crucial for understanding ecosystem interactions.

Purpose of the Study:

• To extend Granger causality for analyzing complex ecosystem time series using machine learning.
• To develop a method that can handle stochastic and nonlinear data inherent in ecosystem dynamics.
• To improve the inference of interaction networks and identify causal drivers in ecosystems.

Main Methods:

• Utilizing an ensemble of artificial neural networks for predictive modeling.
• Employing machine learning to select a minimal set of variables for maximizing prediction.
• Quantifying individual variable contributions to overall predictive performance for relationship decomposition.

Main Results:

• Demonstrated improved interaction network inference in mesocosm and simulated ecosystems.
• Successfully identified drivers of cyanobacteria blooms using a long-term lake monitoring dataset.
• Showcased enhanced forecasting capabilities for ecosystem components.

Conclusions:

• The EcohNet approach effectively extends Granger causality to complex ecological time series.
• This method provides interpretable insights into ecosystem drivers and interactions.
• EcohNet offers a powerful tool for analyzing multivariate time series across various scientific domains.