Deep Learning in Ecology: Data-driven Methods for Ecosystem Analysis

Deep learning algorithms have been widely used on predictive problems, but their capacity for generating scientific knowledge has been largely overlooked. In this project I apply deep learning methods to the problem of missing link prediction in ecological networks, and I investigate methods to extract ecological insight from the trained deep learning algorithms. To achieve this, I collate publicly-available ecological networks to form three new datasets, against which I evaluate novel graph-based neural network methods for missing link prediction. Additionally, I devise two methods for gaining scientific understanding from trained graph-based neural networks, and compare the results to findings in the literature and against previously-used methods. I show that these deep learning approaches outperform traditional machine learning algorithms, and appear to learn useful information about the topology of ecological networks. I conclude that the ability of deep learning algorithms to learn complex patterns in data could help researchers to understand the underlying behaviour of complex systems.