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：近年来，在因果推断框架下生成推荐的研究呈上升趋势，其中推荐被视为一种干预手段。这一方法加深了我们对推荐如何影响用户行为的理解，并有助于识别促成这种影响的因素。许多推荐系统因果推断领域的研究人员关注使用倾向性得分，这可以减少偏差，但也可能引入额外方差。其他研究提出使用来自随机对照试验的无偏数据，尽管这一方法需要某些在实践中可能难以满足的假设。在本文中，我们首先探讨了推荐的因果可解释性，并展示了底层曝光机制可能使观测反馈的最大似然估计产生偏差。鉴于混杂因素可能难以测量，我们提出使用对比自监督学习来减少曝光偏差，具体通过使用逆倾向性得分和扩展正样本集实现。基于理论发现，我们引入了一种新的对比反事实学习方法，该方法整合了三种基于估计曝光概率或随机反事实样本的新型正采样策略。通过在两个真实数据集上的广泛实验，我们证明我们的对比反事实学习方法优于现有最先进方法。