Affective homophily as the dominant organizing principle in online conflict discourse networks.

This study develops a diagnostic framework using a Graph Attention Network (GATv2) to uncover the organizing principles of online conflict networks. By modeling link prediction and rigorously comparing GATv2 against multiple baselines, we validate its superior performance and interpretability. Our analysis delivers a "computational verdict": the model learns to overwhelmingly prioritize affective homophily while systematically discounting users' structural prestige. Probing the learned embeddings reveals that emotion is the dominant dimension of this homophily, a phenomenon we term "emotional fortresses." Critically, by analyzing the model's prediction failures, we identify a novel class of "emergent brokers" who bridge emotional divides through low centrality and affective neutrality. Finally, applying the framework to non-conflict domains empirically establishes the theory's boundary conditions. This research provides a new theoretical lens and methodological tool for understanding network dynamics in polarized environments.