Weak interactions cause poor performance of common network inference models

1. Network inference models have been widely applied in ecological, genetic and social studies to infer unknown interactions. However, little is known about how well the models perform and whether they produce reliable results when confronted with networks where weak interactions predominate and for different amounts of data. This is an important consideration as empirical interaction strengths are commonly skewed towards weaker interactions, which is especially relevant in ecological networks, and a number of studies suggest the importance of weak interactions for ensuring the dynamic stability of a system. 2. Here we investigate four commonly used network methods (Bayesian Networks, Graphical Gaussian Models, L1-regularised regression with the least absolute shrinkage and selection operator, and Sparse Bayesian Regression) and employ network simulations with different interaction strengths to assess their accuracy and reliability. 3. The results show poor performance, in terms of the ability to discriminate between existing relationships and no relationships, in the presence of weak interactions, for all the selected network inference methods. 4. Our findings suggest that though these models have some promise for network inference with networks that consist of medium or strong interactions and larger amounts of data, data with weak interactions does not provide enough information for the models to reliably identify interactions. Therefore, networks inferred from data of that type should be interpreted with caution. ### Competing Interest Statement The authors have declared no competing interest.