Enhancing Advertising Recommendation Performance via Integrated Causal Inference and Exposure Bias Correction

Abstract

 This paper addresses the common problem of exposure bias in advertising recommendation by proposing a new method that integrates causal inference with bias correction to enhance the robustness and accuracy of recommendation systems in complex environments. The study first analyzes the selective exposure characteristics of user behavior data in advertising scenarios and points out that relying only on statistical correlation can easily lead to spurious associations, thereby reducing the reliability of recommendations. On this basis, a causal modeling framework is designed, introducing causal inference tools such as potential outcomes and average treatment effect to model the causal relationship between ad exposure and click behavior. At the same time, inverse propensity weighting is applied to reweight the training data and reduce systematic bias introduced by the exposure mechanism. The method unifies causal structure learning, counterfactual generation, and causal regularization in the modeling process, ensuring that the recommendation results are closer to users' true preferences. The experimental section includes comparative experiments and multi-dimensional sensitivity tests, such as hyperparameter sensitivity, environmental sensitivity, and data sensitivity. Results show that the proposed method achieves superior performance on Precision@10, Recall@10, ACC@10, and NDCG compared with existing methods, and demonstrates high robustness under different experimental conditions. Overall, this study provides a systematic solution for addressing exposure bias and improving recommendation effectiveness in advertising recommendations and offers valuable guidance for future practical applications.