Date of Award
Doctor of Philosophy (PhD)
Graphs are extensively employed in many systems due to their capability to capture the interactions (edges) among data (nodes) in many real-life scenarios. Social networks, biological networks and molecular graphs are some of the domains where data have inherent graph structural information. Built graphs can be used to make predictions in Machine Learning (ML) such as node classifications, link predictions, graph classifications, etc. But, existing ML algorithms hold a core assumption that data instances are independent of each other and hence prevent incorporating graph information into ML. This irregular and variable sized nature of non-Euclidean data makes learning underlying patterns of the graph more sophisticated. One approach is to convert the graph information into a lower dimensional space and use traditional learning methods on the reduced space. Meanwhile, Deep Learning has better performance than ML due to convolutional layers and recurrent layers which consider simple correlations in spatial and temporal data, respectively. This proves the importance of taking data interrelationships into account and Graph Convolutional Networks (GCNs) are inspired by this fact to exploit the structure of graphs to make better inference in both node-centric and graph-centric applications. In this dissertation, the graph based ML prediction is addressed in terms of both node classification and link prediction tasks. At first, GCN is thoroughly studied and compared with other graph embedding methods specific to biological networks. Next, we present several new GCN algorithms to improve the prediction performance related to biomedical networks and medical imaging tasks. A circularRNA (circRNA) and disease association network is modeled for both node classification and link prediction tasks to predict diseases relevant to circRNAs to demonstrate the effectiveness of graph convolutional learning. A GCN based chest X-ray image classification outperforms state-of-the-art transfer learning methods. Next, the graph representation is used to analyze the feature dependencies of data and select an optimal feature subset which respects the original data structure. Finally, the usability of this algorithm is discussed in identifying disease specific genes by exploiting gene-gene interactions.
Bamunu Mudiyanselage, Thosini K., "Machine Learning Methods for Effectively Discovering Complex Relationships in Graph Data." Dissertation, Georgia State University, 2021.
File Upload Confirmation