Loading...
Description
Implicit feedback collaborative filtering recommender systems suffer from exposure bias that corrupts performance and creates filter bubbles and echo chambers. Our study aims to provide a practical method that does not inherit any exposure bias from the data given the information about the user, the choice, and the choice set associated with each observation. We validated the model’s functionality and capability to reduce bias and compared it to baseline mitigation strategies by simulation. Our model inherited little to no bias, while the other approaches failed to mitigate all bias. To the best of our knowledge, we are first to identify a feasible approach to tackle exposure bias in recommender systems that does not require arbitrary parameter choices or large model extensions. With our findings, we encourage the recommender systems community to move away from rating-matrix-based towards discrete-choice-based models.
Recommended Citation
Krause, Thorsten; Stattkus, Daniel; Deriyeva, Alina; Beinke, Jan Heinrich; and Thomas, Oliver, "Beyond the Rating Matrix: Debiasing Implicit Feedback Loops in Collaborative Filtering" (2022). Wirtschaftsinformatik 2022 Proceedings. 12.
https://aisel.aisnet.org/wi2022/ai/ai/12
Beyond the Rating Matrix: Debiasing Implicit Feedback Loops in Collaborative Filtering
Implicit feedback collaborative filtering recommender systems suffer from exposure bias that corrupts performance and creates filter bubbles and echo chambers. Our study aims to provide a practical method that does not inherit any exposure bias from the data given the information about the user, the choice, and the choice set associated with each observation. We validated the model’s functionality and capability to reduce bias and compared it to baseline mitigation strategies by simulation. Our model inherited little to no bias, while the other approaches failed to mitigate all bias. To the best of our knowledge, we are first to identify a feasible approach to tackle exposure bias in recommender systems that does not require arbitrary parameter choices or large model extensions. With our findings, we encourage the recommender systems community to move away from rating-matrix-based towards discrete-choice-based models.