<p>Chronic knee osteoarthritis (OA) pain involves widespread reorganization of brain network topology that often eludes conventional region-of‐interest analyses. Resting‐state fMRI connectivity offers a promising window into these alterations, yet methods that both capture signed (positive vs. negative) interactions and remain interpretable are lacking. We present DualEdge-GAT, an explainable graph neural network that constructs dynamic sliding-window partial-correlation graphs across 116 AAL regions and employs dual-stream graph attention (positive vs. negative edges) to separately weight positively and negatively signed functional connections. The architecture integrates Squeeze-Excitation channel reweighting and Self-Attention Graph Pooling to focus on the most informative nodes and edges before final classification. Evaluated on 56 chronic knee OA patients versus 20 healthy controls, DualEdge-GAT achieved a mean validation accuracy of 87 ± 4% and mean F1 score 92 ± 2%. Attention-based explanations consistently identified posterior cerebellar lobules as the most critical hubs. Graph-theoretic follow-up revealed that these regions exhibit significantly reduced nodal degree in OA patients, indicating a breakdown of richly interconnected local neighborhoods. Against a kernel SVM trained solely on handcrafted nodal degree features, DualEdge-GAT demonstrated superior discriminative performance and delivered mechanistic insight by pinpointing the specific cerebellar circuits underlying chronic pain. DualEdge-GAT not only achieves high accuracy in classifying chronic knee OA pain but also provides a transparent map of circuit-level dysfunction, uncovering cerebellar hub breakdown as a potential biomarker and therapeutic target.</p>