Network structure indicators predict ecological robustness in food webs

Food web robustness is a critical aspect of ecosystem stability and has been extensively studied in ecology. However, the potential of machine learning techniques in predicting food web robustness and the identification of key network structure indicators have not been fully explored. We compared the suitability of different machine learning methods and assessed the relative importance of network structure indicators for predicting the robustness of food webs. We utilized a variety of food web datasets spanning different ecosystems to calculate network structure indicators, which include average distance (AD), betweenness centrality (BC), directional connectivity (C), closeness centrality (CC), diameter (D), degree centrality (DC), edge betweenness centrality (EBC), number of links (L), linkage density (LD), and number of nodes (N). We then compared the performance of machine learning methods, including artificial neural network (ANN), random forest (RF), least absolute shrinkage and selection operator (LASSO), and decision tree (DT), and evaluated the relative importance of network structure indicators on robustness predictions. The results demonstrate that the RF model has the best performance (MAE = 0.0178, RMSE = 0.0263, R² = 0.9063). Meanwhile, the CC indicator has a significant impact in predicting robustness of food webs. It is suggested that both the RF model and the CC indicator should be considered seriously in predicting food web robustness. This research elucidates the differential outcomes when various machine learning methodologies and indicators are employed to predict the robustness of food webs. It significantly enhances our understanding by demonstrating the precise capability of machine learning models in forecasting the robustness of food webs.