The School of Computer Science is pleased to present…
Date: Tuesday, April 8, 2025
Time: 10:30 AM
Location: Essex Hall, Room 122
Social networks are diverse interactions that can generally be classified as positive (e.g., friendships, likes) or negative (e.g., conflict, dislikes). While link prediction—anticipating the formation of new ties—has been widely explored, the task of predicting link breakups, where existing ties weaken or dissolve, remains relatively understudied. These transitions are often subtle and gradual, frequently going undetected until negative outcomes have already occurred. Current systems rely heavily on user manual intervention to flag such changes, making them both inefficient and reactive. In this study, we address the challenge of automatically predicting potential link breakups by analyzing both structural and behavioral cues within social network graphs. Building upon the ClassReg heuristic, we introduce HeuGAT, an enhanced approach that replaces the original deep learning layer with a Graph Attention Network (GAT) layer. This integration allows the model to more effectively capture the contextual significance of neighboring nodes through attention mechanisms.
Internal Reader: Dr. Dan Wu
External Reader: Dr. Sudhir Paul
Advisor: Dr. Ziad Kobti
