@InProceedings{10.1007/978-3-031-78541-2_2,
author="Farokhi, Soheila
and Foumani, Arash Azizian
and Qi, Xiaojun
and Derr, Tyler
and Karimi, Hamid",
title="EDGE-UP: Enhanced Dynamic GNN Ensemble for Unfollow Prediction in Online Social Networks",
booktitle="Social Networks Analysis and Mining",
year="2025",
publisher="Springer Nature Switzerland",
address="Cham",
pages="20--39",
abstract="In the complex landscape of online social networks, predicting unfollow events is challenging due to data sparsity, class imbalance, and the dynamic nature of user interactions. This paper presents EDGE-UP, an Enhanced Dynamic Graph Neural Network (GNN) Ensemble model adeptly designed to overcome these challenges in unfollow prediction. EDGE-UP leverages a large-scale, longitudinal Twitter dataset featuring 58 weekly snapshots across 118,890 users to capture the evolving social dynamics. It minimizes the need for extensive feature engineering by utilizing GNNs for spatial encoding and LSTMs for capturing temporal dynamics, addressing data sparsity and class imbalance through ensemble learning and negative sampling strategies. Our experiments demonstrate EDGE-UP's superior performance in accurately predicting unfollow events, setting a new standard in social network analysis, and offering versatile applicability across different platforms. The code and data are available here: https://github.com/DSAatUSU/edge-up.",
isbn="978-3-031-78541-2"
}