Graph Neural Network with Quasi - Data Augmentation for Modelling Food Web Relationships

Several of the United Nations sustainable development goals (SDGs) are related via the food web. However, most intervention measures have been designed to achieve specific goals in isolation without assessing its impact on other goals. The food web in nature's ecosystem represents a network that can potentially serve as a proxy towards assessing the impact of certain goals on others via modern computational modelling. While previous works have largely focused on carrying out several types of analysis using graph inspired methods, none has focused on predicting the nature of relationships on the food web. In this work therefore, we propose a graph-based approach leveraging recent advances in graph neural network that captures the interactions and feeding relationships in nature's ecosystem. More specifically, we focus on two tasks of predicting first, if there exist relationships between several entities in the food web and second, the nature of these relationships given there exist one. As with all real-world datasets for sustainability research, the food web datasets are limited in size which impacts generalization negatively. To bypass this problem, we propose a novel data augmentation method. We evaluate the performance of the proposed approach using publicly available integrated food web datasets from three estuaries and also compare with multiple baselines. Our results demonstrate the effectiveness of the proposed approach, outperforming competing baselines on a variety of experiments. We also use a network analysis approach to identify the keystone species in this food web. Overall, our work has great implications mostly in the area of ecosystem monitoring and restoration.