Promoting Explainability and Unbiasedness in Ranking-Based Recommendation

What is it about?

Recent research has questioned the fairness of recommender systems, with a particular interest in addressing bias and lack of transparency in black box models. In this paper, we make significant improvements to Bayesian Personalized Ranking (BPR), considered to be a state of the art ranking-based recommender system model. First we add explainability to BPR. Second, we address the issue of exposure bias, which stems from the assumption that items that have not been exposed to users are not relevant. This exposure bias usually translates into an unfairness against the least popular items because they risk being hidden by the recommender system. As a result, our proposed ranking-based recommender system models aptly capture both debiased and explainable user preferences.

Featured Image

Photo by Alexander Hafemann on Unsplash

Photo by Alexander Hafemann on Unsplash

Why is it important?

The lack of explainability and exposure bias are both notorious problems in recommender systems. Improving the explainability of a recommender system can help the user scrutinize the predictions for potential bias or errors. Explanations can also help users make more informed decisions based on the algorithmic recommendations that they receive, such as whether to follow a recommendation. Exposure bias usually results in the recommender system being unable to accurately learn the true preferences of the user. For instance, items that are relevant to a user’s interests can be considered to be non relevant just because the user did not have a chance to see them or interact with them yet. Exposure bias can lead to less popular items being hidden from the user or under-exposed because a biased recommender system cannot learn to model their relevance. In other words the recommender system is biased by initial under-exposure and this bias results in even more under-exposure of these items, creating a bias feedback loop.