Date of Award
Master of Science (MS)
Electrical Engineering and Computer Science
Graph Augmentation, Graph Representational Learning, Spectral Moments
Graph representational learning focuses on learning real value vectors that for nodes,edges or the graph, such that these vectors capture adequate information about these entities. Graph data augmentation, focuses on changing the structure or features in a graph to help improve classification performance and become more generalizable. This can be broadly categorized into feature based augmentation and structure based augmentation. Feature augmentation focuses on changing the feature matrix, without changing the structure of the graph to help improve the performance of the graph neural network. Graph structure augmentation refers to the manipulation of the adjacency matrix of a given graph to achieve better classification performance.
Our approach focuses on the problem of graph augmentation but from a spectral standpoint. More specifically, we attempt to augment a graph using spectral moments. Recent results have indicated that the second, third and fourth spectral moments of a graph, have strong connections to the graph's properties, such as degree distribution, clustering coefficient, and connectivity. Our contribution is two fold: First, we explain a formal method to find a spectral moment that helps maximize node classification performance. Second, we also provide an algorithm to augment the graph using it's spectral moments, and therefore augment the graph to the spectral point that helps maximize classification performance while making the graph sparse. For the purpose of node classification, we use the GraphSAGE model with no node sampling and the mean aggregator. We notice that the node classification performance after augmentation goes up in a majority of our datasets, and furthermore, the graph also gets sparser across all our datasets.
Gopalakrishnan, Atul Anand, "Graph Augmentation using Spectral moments" (2023). Theses - ALL. 701.