The Application of Causal Inference Algorithms in Federated Recommender Systems

Abstract

This study introduces the application of causal inference algorithms in federated recommender systems through a framework called FedCIRec. The framework utilizes two key causal inference methods: Instrumental Variable (IV) and Counterfactual Inference (CF). IV helps identify and address confounding biases, improving recommendation accuracy. CF addresses data loss issues in federated learning, enabling the inference of possible scenarios and outcomes. Experimental results on the MIND and Adressa datasets demonstrate excellent performance in evaluation metrics. The FedCIRec framework has advantages over centralized storage and user privacy protection methods. Further exploration of causal inference methods and optimization of recommendation effects while protecting user privacy are areas for future research. Introducing causal inference techniques in federated recommendation systems has broad research and practical application prospects.