Computational methods for identifying structure in ecological networks

Background/Question/Methods

Many ecological processes are distributed over networks. Recent studies have shown that the geometry of plant-pollinator, predator-prey, and competitive networks is central to their stability and robustness. Identifying the links among ecological actors from observational data is, however, a largely open problem. One strategy is to exploit the sparsity of the envisioned solution. Recent developments in machine learning have shown that L1 regularization may often be used to find stable sparse solutions to a variety of statistical problems. To operationalize this approach and study its performance, I simulated medium-sized communities of species interacting according to a Generalized Lotka-Volterra model in a randomly fluctuating environment. Such systems are known to have a stationary distribution with inverse covariance structure reflecting the network of directly interacting species. I then used the glasso algorithm for L1 estimation of graphical models to study the accuracy with which interacting species could be determined using only data from the stationary time series (i.e., without press/pulse perturbations). This problem was previously believed to be intractable.

Results/Conclusions

In communities of 10 to 100 interacting organisms, the glasso algorithm classified species pairs as interacting or non-interacting with an average accuracy of 75%. Overall accuracy was insensitive to the number of interacting species, average interaction strength, and the variance of environmental fluctuations. Accuracy declined with increasing network connectance. Sensitivity (correctly classifying interacting species pairs), similarly, was insensitive to the number of interacting species, average interaction strength, and the variance of environmental fluctuations and declined with connectance. Specificity (correctly classifying non-interacting species pairs) was relatively insensitive to all manipulated variables. These results show that automated approaches to detecting species interactions are feasible with current algorithms. I suggest that further advances might be obtained by combining L1 regularization with autoregressive strategies to exploit temporal dependence in time series data.