Causal inference for alleviating confounding bias in multi-criteria rating recommendation

Abstract

Integrating multi-criteria (MC) ratings into recommender systems can enhance the service quality of online platforms. MC ratings depict more fine-grained user preferences from multiple dimensions, such as a hotel system, including ratings for overall, location, cleanliness, etc. The existing MC methods focus on mining the correlation from historical interactions through the data-driven paradigm. However, the traditional methods may capture spurious association in biased observations due to various confounders, which can reduce prediction accuracy. So far, research on how to alleviate confounding bias in MC rating recommendation scenarios remains unexplored. To fill this research gap, we propose a novel Deconfounding Multi-Criteria Recommendation (DMCR) framework, which is used to mitigate the harmful impact triggered by confounders. Specifically, we block the back-door paths that cause bias through the front-door adjustment and estimate the causal effect between user-item pair and overall rating. In the inference phase, the DMCR approximates the outcome after intervention by conditional probabilities on the observational MC data. Moreover, we leverage graph neural network to model underlying higher-order dependencies in MC ratings. This modeling scheme helps to develop the heterogeneity of user MC behavioral preferences. Experimental results on six real datasets demonstrate that the DMCR outperforms the existing baselines.