Recommender systems have become an indispensable part of web systems. They prevent their users from being overloaded with a flood of information by suggesting relevant items to them based on prior user interactions. However, recommender systems frequently face popularity bias issues: popular items are over-represented in the recommendations while less popular ones get less exposure. Consequently, the popularity bias not only threatens the fairness of recommendation results but also decreases the degree of personalisation of recommendations as popularity is not a direct indicator of personal relevance. Since the feedback loop further amplifies the popularity bias, debiasing strategies need to be developed to provide a fair and diverse recommender system.
The proposed approach identifies and reduces the popularity bias based on a predefined fairness policy. To do this, we modify PageRank, a graph-based algorithm developed by Google Search, which finds relevant users in a network. Fairness is integrated locally by adjusting the transition probabilities of a random walk on a graph according to the fairness policy. Depending on the desired fairness goal measured by the generalised Gini index, the results of this work present suitable fairness policies. In order to evaluate the method, experimental results are presented comparing the applied modifications to the unconstrained PageRank. The experiments are based on a public data set from the online social network Twitter.