Machines learn ecological networks: automated discovery of ecological networks based on empirical data

Constructing ecological networks is known to be important and challenging in community ecology. In particular, to construct the holistic structure of ecological networks, identifying species interaction is essential but often costly and impalpable. Recent studies providing major challenges in assembling ecological networks have highlighted the need of new and more powerful approaches to reconstruct biological networks, including species interaction networks. In literature, there are no promising verifications in using machine leaning (ML) approaches to reconstruct ecological networks. In this work, we develop and employ a variety of ML methods, including penalized regression and graphical tools, to reconstruct ecological networks. For evaluation, we apply the methods to empirical time series data sets of 20 species abundances collected at Lake Constance in central Europe. We use resampled data to identify highly-ranked interactions among species and measure their consistency across 7 ML methods and 5,000 learning processes. We show that the best precision, recall, and F1 score were 0.48, 0.97, and 0.64, respectively, among all penalized regression methods under comparison. In summary, our study shows that machine learning methods offer promising data-driven and automated tools for reconstructing ecological networks and discovering underlying biological interactions among species.