CATS: Combinatorial Allocation via Thompson Sampling for Multi-Buyer–Multi-Seller Data Markets

Modern machine-learning systems rely on valuable data distributed across multiple sellers, while different buyers may derive heterogeneous predictive value from the same data. We study the sequential allocation of sellers’ data among competing buyers when buyer–seller values are initially unknown and allocations must be exclusive. We formulate this problem as a combinatorial semi-bandit and propose CATS, which uses Thompson sampling to learn buyer-specific data values while selecting welfare-maximizing feasible allocations. On a real-world credit-scoring dataset, CATS reduces cumulative regret by 93.3% compared with a combinatorial upper-confidence-bound benchmark. In the contextual setting, contextual CATS improves welfare by 8% over its static counterpart and approaches contextual-oracle performance.