De-biasing user preference ratings in recommender systems

Prior research has shown that online recommendations have significant influence on users' preference ratings and economic behavior. Specifically, the self-reported preference rating (for a specific consumed item) that is submitted by a user to a recommender system can be affected (i.e., distorted) by the previously observed system's recommendation. As a result, anchoring (or anchoring-like) biases reflected in user ratings not only provide a distorted view of user preferences but also contaminate inputs of recommender systems, leading to decreased quality of future recommendations. This research explores two approaches to removing anchoring biases from self-reported consumer ratings. The first proposed approach is based on a computational post-hoc de-biasing algorithm that systematically adjusts the user-submitted ratings that are known to be biased. The second approach is a user-interface-driven solution that tries to minimize anchoring biases at rating collection time. Our empirical investigation explicitly demonstrates the impact of biased vs. unbiased ratings on recommender systems' predictive performance. It also indicates that the post-hoc algorithmic debiasing approach is very problematic, most likely due to the fact that the anchoring effects can manifest themselves very differently for different users and items. This further emphasizes the importance of proactively avoiding anchoring biases at the time of rating collection. Further, through laboratory experiments, we demonstrate that certain interface designs of recommender systems are more advantageous than others in effectively reducing anchoring biases.