The School of Computer Science is pleased to present…
Heuristic and Graph Attention Network (HeuGAT): a GNN layer substitution for the DL layer in the ClassReg heuristic to enhance link and breakups prediction in social network structures
MSc Thesis Defense by: Hridoy Pal
Date: Friday, June 06, 2025
Time: 10:00 AM
Location: Memorial Hall, Room 109
Social networks consist of a wide range of interactions, which are generally categorized as either positive, such as friendships or likes, or negative, such as conflicts or dislikes. While the prediction of new links, commonly known as link prediction, has been widely studied in the field of social network analysis, the task of predicting link breakups, where existing relationships weaken or disappear, has received relatively little attention. These changes are often gradual and subtle, making them difficult to identify before any negative consequences take place. Most existing systems depend largely on manual input from users to recognize and report such changes, making the overall process both inefficient and reactive. In this research, we address the challenge of automatically identifying potential link breakups by analyzing both the structural and behavioural patterns present in social network graphs. Building upon the ClassReg heuristic framework, we propose HeuGAT, an improved model that replaces the original deep learning component with a Graph Attention Network. This modification enables the model to better capture the contextual relevance of neighbouring nodes by applying attention-based learning within the social network graph.
Internal Reader: Dr. Dan Wu
External Reader: Dr. Sudhir Paul
Advisor: Dr. Ziad Kobti
Chair: Dr. Muhammad Asaduzzaman
