Fig. 4

Ensemble method provides robust parameterization of generalized Lotka–Volterra (gLV) network models. a Probability distribution of the symmetric mean absolute percentage error (sMAPE) across 1000 gLV models perturbed from the best-fit model given by latent gradient regression (LGR). The distribution has three peaks that partition the 1000 models into three groups that represent good (668 models), mediocre (53 models) and poor (279) fits to data. Models in the first group were combined to make an ensemble. Dashed purple line: unperturbed model. b Simulated trajectories of the fish abundance data by models from the three groups. Unperturbed model is the best-fit model given by LGR. c The coefficient of variation (CV) of gLV parameters across the 668 models in the ensemble