Modeling item exposure and user satisfaction for debiased recommendation with causal inference.

Recommender systems (RSs) aim to provide suggestions for items that are most pertinent to a particular user. Typically, RSs are trained and evaluated directly on the observed items, raising concerns about exposure bias - many missing items are false negatives, which were not consumed due to lack of exposure rather than lack of affinity. In addition, user satisfaction is often ignored in previous RSs, where consumer behaviors may be influenced by advertisement or promotion instead of actual user interests. In this paper, we propose a novel model-agnostic causal inference method for debiased recommendation, which models item E xposure and user S atisfaction simultaneously with C ausal I nference (ESCI). Specifically, we formulate a causal graph to describe the recommendation process, where the ranking score is influenced by item exposure, user satisfaction, and user-item matching. We investigate the change in the ranking score when item exposure is discarded. In addition, we propose an adversarial training strategy to improve the generalization and robustness of recommender systems. During testing, we perform causal inference to remove the effect of item exposure. The comprehensive experimental study on four benchmark datasets demonstrates that the proposed ESCI enhances recommendation performance for users with non-high interaction frequencies, thereby outperforming state-of-the-art baselines. [ABSTRACT FROM AUTHOR]