This paper proposes and evaluates two algorithms for recommendation systems. Most recommendation systems concentrate on optimizing one primary metric, for example, the advantage of a consumer purchasing an item or minimizing the cost of operations. Abdollahpouri argues that the benefits of other stakeholders should also be considered in such recommendation systems. Examples of other stakeholders include, but are not limited to, other users of the system, the providers of the products, and even the recommendation systems themselves.
The research tries to establish a common, more general framework to study recommendation systems with multiple objectives. In the proposed framework, the objectives are divided into two classes, local and global. The local objectives are for individual users; the global objectives are for other users or the systems. These objectives are further categorized as user related, item related, or user-item related.
The author proposes two algorithms. The first one is fairness-aware regularization, which identifies “a regularization component of the objective function that will be minimized when the distribution of recommendations is fair in terms of popular and non-popular items.” The second one is post processing re-ranking, “which explicitly accounts for the various aspects associated with an under-specified query.”
The paper compares the results of the two algorithms against the base algorithm of RankALS, “a pair-wise learning-to-rank algorithm.” The evaluation uses the Epinions dataset, “which is gathered from a consumer opinion site where users can review items.” This dataset consists of “664,824 ratings given by 40,163 users to 139,736 items.” The results show that the proposed algorithms perform better than the baseline algorithm.
The proposed algorithms are effective according to the presented results. The paper is short due to limited space, so some details and results are missing. However, the included ideas and algorithms are interesting. Furthermore, the proposed ideas should help researchers in related areas explore this particular direction, as well as give practitioners some evidence that the algorithms should work well since the test datasets are reasonably large.