Ethics in AI: Collaborative filtering based approach to alleviate strong user biases and prejudices

Generally, recommender systems tend to satisfy users' tastes and preferences by showing them the extreme versions of the content they like. This over time results in reinforcement of user's biases, which has far reaching socioeconomic and political consequences. For example, Facebook played a major part in inciting the genocide against Rohingya minority in Myanmar in 2017 by reinforcing users preferences for content containing hate speech to maximize impressions and clicks. This paper proposes to blend user's recommendations with items which are of “high quality” but are inversely correlated to user's taste, which might help in loosening up user's biases or at least help prevent heavy reinforcement of user's biases. A collaborative filtering based algorithm is proposed to find “high quality” items which are inversely correlated to user's tastes. The proposed algorithm is implemented and tested on the MovieLens 1M dataset and the results are compared with that of a traditional collaborative filtering approach on the basis of two metrics - quality and diversity of recommendations for each user. The proposed algorithm shows better performance in both metrics.