A Theoretically‐Grounded Federated Attribution Framework with Adaptive Differential Privacy Budgets for Cross‐Device Social Commerce Advertising Systems

In social electronic commerce advertising systems involving multiple devices, user behavior data is scattered across various devices and is highly sensitive regarding privacy. How to achieve ad attribution of high quality while protecting user privacy has become a key issue. This paper proposes a theoretically supported Federated Attribution Framework, which innovates on the basis of existing Shapley value attribution methods and federated learning mechanisms. By integrating user behavior graph modeling across multiple devices, introducing graph neural networks for local temporal encoding, and implementing a federated alignment mechanism consisting of two stages, it achieves collaborative user representation and attribution optimization across devices. Additionally, an adaptive differential privacy budget allocation strategy is proposed, which can dynamically adjust privacy budget allocation based on device attribution sensitivity and training rounds, achieving a personalized balance between privacy protection and attribution performance. Experimental results show that the proposed method improves attribution accuracy by an average of 8.3% on a social electronic commerce advertising dataset compared to existing methods.