User-item matching for recommendation fairness: a view from item-providers
share
As we all know, users and item-providers are two main groups of participants in recommender systems. The main task of this paper is to significantly improve the coverage fairness (item-provider oriented objective), and simultaneously keep the recommendation accuracy in a high level (user oriented objective). First, an effective and totally robust approach of improving the coverage fairness is proposed, that is to constrain the allowed recommendation times of an item to be proportional to the frequency of its being purchased in the past. Second, in this constrained recommendation scenario, a serial of heuristic strategies of user-item matching priority are proposed to minimize the loss of recommendation accuracy. The parameterized strategy among them is validated to achieve better recommendation accuracy than the baseline algorithm in regular recommendation scenario, and it has an overwhelming superiority in coverage fairness over the regular algorithm. Third, to get the optimal solution of this user-item matching problem, we design a Minimum Cost Maximum Flow model, which achieves almost the same value of coverage fairness and even better accuracy performance than the parameterized heuristic strategy. Finally, we empirically demonstrate that, even compared with several state-of-the-art enhanced versions of the baseline algorithm, our framework of the constrained recommendation scenario coupled with the MCMF user-item matching priority strategy still has a several-to-one advantage in the coverage fairness, while its recommendation precision is more than 90 What is more, our proposed framework is parameter-free and thus achieves this superior performance without the time cost of parameter optimization, while all the above existing enhanced algorithms have to traverse their intrinsic parameter to get the best performance.
READ FULL TEXT