<p>Predicting quality of experience (QoE) in dynamic edge networks is a critical yet challenging task due to the high-entropy noise and the complex coupling between spatial topology and temporal traffic patterns. Existing architectures often suffer from ‘noise-blindness,’ failing to capture spatiotemporal dependencies or yielding sub-optimal generalizations in volatile environments. To overcome these barriers, we introduce GATFormer-MultiTask, a hybrid deep learning framework specifically engineered for robust edge intelligence. The model integrates graph attention networks (GAT) for topology-aware spatial feature extraction with Transformer encoders to model long-range temporal evolution. Crucially, we deviate from traditional single-objective modeling by implementing a multi-task learning paradigm that jointly optimizes for QoE regression and discrete classification. This dual-objective strategy serves as a high-level regularization mechanism, stabilizing the latent feature space against stochastic network jitter. Empirical validation on the real-world DATA7 dataset demonstrates that GATFormer-MultiTask significantly outperforms state-of-the-art baselines, achieving a 45% reduction in RMSE (from 0.0227 to 0.0125). Notably, the proposed framework achieves a positive R² score of 0.0101, effectively reversing the negative predictive trends observed in single-task models. These results substantiate that the synergy of structural-temporal encoding and multi-faceted optimization provides a superior, noise-resilient solution for real-time edge diagnostics.</p>