Inductive link prediction boosts data availability and enables cross-community link prediction in ecological networks

Barry Biton, Rami Puzis, Shai Pilosof 

Predicting species interactions (links) within ecological networks is crucial for advancing our understanding of ecosystem functioning and responses of communities to environmental changes. Traditional link prediction models are often constrained by sparse, incomplete data and are typically limited to single networks. Here, we address these issues using an innovative inductive link prediction (ILP) approach. By pooling data across communities and applying transfer learning, our model predicts interactions within and between ecological networks. 
We evaluated the model performance on 538 networks across four community types: plant-seed disperser, plant-pollinator, host-parasite, and plant-herbivore. ILP models outperform non-ILP models, achieving a mean balanced accuracy of 0.68 compared to 0.57. Furthermore, ILP consistently predicts missing links more effectively across fractions of missing links. Despite its strengths, cross-community prediction efficacy varies, with plant-seed disperser and host-parasite networks showing better performance than plant-pollinator and plant-herbivore networks as both training and test sets. 
Finally, leveraging ILP’s generalizability, we developed a pre-trained model that ecologists can readily use to make instant predictions for their networks. This study highlights ILP’s potential to improve prediction of ecological interactions, enabling generalization across diverse ecological contexts and bridging critical data gaps.