Key points are not available for this paper at this time.
The application of federated edge learning to cross-domain multi-task learning—where clients hold data from different domains and tasks—introduces significant challenges, specifically the dual skew where the intersection of feature and task heterogeneity creates a vicious circle of internal error propagation. Although prior researches have advanced optimization strategies, they often relies on static parameter isolation or monolithic aggregation, neglecting the potential synergistic relationships among tasks, thereby leading to optimization deadlocks. To bridge this gap, we propose CRAFT, a novel method designed with a two-timescale optimization philosophy that integrates Cross-staggered learning, a coalition foRmation gAme, and selF-adapTive weighting. Specifically, CRAFT mitigates local empirical risk arising from dual skew through a cross-stagger paradigm, which employs a dynamic, alternating optimization strategy to decouple conflicting gradients. It corrects offline gradient bias by modeling item synergy via a coalition formation game for structural stability, while resolving online update conflicts using a causal inference-based mechanism that estimates the true counterfactual contribution of each client to filter real-time noise. Extensive experiments under a challenging unbalanced cross-domain hybrid-task setting demonstrate that CRAFT consistently outperforms existing approaches, achieving up to an improvement of 7.26% in converged accuracy with superior robustness across varying heterogeneity levels.
Xiang et al. (Sun,) studied this question.