In this paper we demonstrate that machine learning (using Abductive ILP) can generate plausible and testable food webs from ecological data. In this approach, unlike previous applications of Abductive ILP, the abductive predicate ‘eats’ is entirely undefined before the start of the learning. We also explore a new approach, called Hypothesis Frequency Estimation (HFE), for estimating probabilities for hypothetical ‘eats’ facts based on their frequency of occurrence when randomly sampling the hypothesis space. The results of cross-validation tests suggest that the trophic networks with probabilities have higher predictive accuracies compared to the networks without probabilities. The proposed trophic networks have been examined by domain experts and comparison with the literature shows that many of the links are corroborated by the literature. In particular, links ascribed with high frequency are shown to correspond well with those having multiple references in the literature. In some cases novel high frequency links are suggested, which could be tested.