Imperfect Detection in Heterogeneous Complex Ecological Network: A GNN-based Approach

Link prediction is challenging for complex ecological networks as the observed data are often incomplete during field sampling. Given a plant-pollinator network, models using a non-negative matrix factorisation and Poisson N-mixture successfully address the ecological issues with imperfect detection while establishing a binary model for predicting unobserved links. We drive an advanced nonlinear approach using a neural network while exploiting a hybrid strategy of the Poisson N-mixture model and the Gaussian mixture model in statistical ecology. This generates a predictive model for link prediction along with interaction probabilities. Evaluation of our method on test data revealed an AUROC (area under the receiver operating characteristic curve) of 85.4% and AUPRC (area under the precision-recall curve) of 80.1% which is significantly higher compared to the current state-of-the-art solutions. Moreover, it has theoretical support since its optimization technique ensures convergence and scalability. (Code: https://github.com/psil123/EcologyGNN)