P-MMF: Provider Max-min Fairness Re-ranking in Recommender System

FairSync: Ensuring Amortized Group Exposure in Distributed Recommendation Retrieval

“…In recent times, the need for fair recommender systems (RS) has gained prominence in industrial requirements [21,33]. Among these requirements, RS platform has the demands of guaranteeing that specific groups achieve a minimum level of exposure to items within a defined period, aligning with the perspective of amortized group max-min fairness (MMF) [2,18,25,38,41]. For instance, certain studies propose to ensure minimum item exposures for new providers [4,28,38,39] for attracting providers to join, while others focus on enhancing the visibility of specific item categories [45] for promoting certain festivals.…”

“…Among these requirements, RS platform has the demands of guaranteeing that specific groups achieve a minimum level of exposure to items within a defined period, aligning with the perspective of amortized group max-min fairness (MMF) [2,18,25,38,41]. For instance, certain studies propose to ensure minimum item exposures for new providers [4,28,38,39] for attracting providers to join, while others focus on enhancing the visibility of specific item categories [45] for promoting certain festivals. Such "minimum wage policy" [32] significantly contributes to the enhancement of RS, fostering the creation of a more equitable and robust ecosystem.…”

“…In modern RS, two-stage pipelines have been widely adopted, as shown in Figure 1 (a). The primary objective of stage-1 (retrieval) is to efficiently generate a small set of candidates from millions of items in a distributed manner within milliseconds [5,8,20,23] while stage-2 (ranking) more accurately deals with the candidates selected in stage-1 and generates the final recommendations (usually single-digit items) [38,45]. Regarding ensuring minimum group exposures in RS, most existing studies [4,9,26,28,37,38,45] primarily concentrated on stage-2.…”

“…Though critically important, existing approaches designed for stage-2 [9,26,28,37,38] cannot be directly applied to stage-1 because they ensure the amortized group exposures by traversing all items and adjusting exposures through aggregating information. During stage-1, however, traversing all items (usually millions of items) is infeasible because (1) these items are distributively stored at different servers, and (2) accessing millions of items causes substantial computational cost.…”

“…In particular, the RS platform first sets a target to ensure that every group attains a minimum level of exposure. Then we approach the problem by formulating it as a distributed resource allocation problem [1,38], with the constraint of required exposures. Subsequently, we can transform the constrained optimization problem into an unconstrained dual problem.…”