Access to this thesis is limited to Boise State University students and employees or persons using Boise State University facilities.

Off-campus Boise State University users: To download Boise State University access-only theses/dissertations, please select the "Off-Campus Download" button and enter your Boise State username and password when prompted.

Publication Date

5-2024

Date of Final Oral Examination (Defense)

3-15-2024

Type of Culminating Activity

Thesis - Boise State University Access Only

Degree Title

Master of Science in Computer Science

Department Filter

Computer Science

Department

Computer Science

Supervisory Committee Chair

Edoardo Serra, Ph.D.

Supervisory Committee Member

Francesca Spezzano, Ph.D.

Supervisory Committee Member

Marion Scheepers, Ph.D.

Abstract

Graph-structured data has increasing applicability in hundreds of domains, the majority of which produce mostly unlabeled data. For this reason, the supervised methods of graph representation learning, which dominate current methods, have limited use in many cases. Similarly, many methods rely on a graph being homophilic in which proximity implies similarity. In many applications, the structure of a node’s connections carries more relevant information than the nodes to which it is connected. The proposed method is an unsupervised graph neural network that doesn't rely on homophily of the graphs. It creates an identifying signature for each node and supplements each node's features with the IDs of its neighbors to encourage representations that can be used to reconstruct the features of the neighbors. The produced representations may be used in downstream tasks such as node classification. Experiments show that ID-GNN is within 3% of the node classification accuracy of state-of-the-art methods on heterophilic graphs.

DOI

https://doi.org/10.18122/td.2186.boisestate

Download
Off-Campus Download

Share

COinS