<p>In recent years, large-scale models have excelled in many machine learning tasks, including the generation of distributed spatial data, but their high training and inference costs pose significant challenges. To address this, researchers have turned to sparsely activated models like Mixture of Experts (MoE), which reduce computational overhead by dynamically activating only a subset of experts. However, MoE’s top-k routing often leads to load imbalances, causing performance bottlenecks during distributed spatial data generation. This work models MoE’s top-k routing as a combinatorial multi-armed bandit (CMAB) problem with a fairness constraint for balanced spatial data generation. In this sense, we propose FairCMAB, a novel algorithm that extends the Upper Confidence Bound (UCB) approach to manage the exploitation-exploration trade-off and integrates a lottery mechanism to ensure fair expert selection while maintaining spatial data quality. Our comprehensive theoretical analyses and evaluations demonstrate that FairCMAB effectively balances expert load within a bounded time step, leading to efficient utilization and improved generation performance for distributed spatial data. To the best of our knowledge, this is the first work to address MoE’s top-k routing as a CMAB problem with fairness for distributed spatial data generation, offering a significant step towards optimizing the efficiency and quality of spatial data generation in practical applications.</p>