There has been a recent surge in the study of generating recommendations within the framework of causal inference, with the recommendation being treated as a treatment. This approach enhances our understanding of how recommendations influence user behaviour and allows for identification of the factors that contribute to this impact. Many researchers in the field of causal inference for recommender systems have focused on using propensity scores, which can reduce bias but may also introduce additional variance. Other studies have proposed the use of unbiased data from randomized controlled trials, though this approach requires certain assumptions that may be difficult to satisfy in practice. In this paper, we first explore the causality-aware interpretation of recommendations and show that the underlying exposure mechanism can bias the maximum likelihood estimation (MLE) of observational feedback. Given that confounders may be inaccessible for measurement, we propose using contrastive SSL to reduce exposure bias, specifically through the use of inverse propensity scores and the expansion of the positive sample set. Based on theoretical findings, we introduce a new contrastive counterfactual learning method (CCL) that integrates three novel positive sampling strategies based on estimated exposure probability or random counterfactual samples. Through extensive experiments on two real-world datasets, we demonstrate that our CCL outperforms the state-of-the-art methods.

翻译：暂无翻译