1. Abstract
Recommender system has been deployed in a large amount of real-world applications, profoundly influencing people's daily life and production. Traditional recommender models mostly collect as much as possible user information to accurate estimate the user preference. However, in real-world scenarios, the users may not want all their behaviors to join into the model training process. For example, the user may want to actively edit her profile by removing the items which are incorrectly clicked or purchased for the other people. In this paper, we study a novel recommendation paradigm, where the users are allowed to indicate their ``willingness'' on letting different data to train the model, and the models are optimized to maximize the utility which trades-off the recommendation performance and the violation of the user ``willingness''. More specifically, we formulate the recommendation problem as a multiplayer game. Each user is a player, and the player action is a selection vector representing whether the user would like to leverage her interacted items to train the model. For efficiently solving this game, we design an influence function based model, which can approximate the recommendation performances for different actions without re-optimizing the model. In addition, we also improve the above model by deploying multiple anchor actions for the influence function, which is expected to improve the performance approximation accuracy. At last, we theoretically analyze the convergence rate of our algorithm and demonstrate the superiority of introducing multiple anchor actions. We conduct extensive experiments based on both simulation and real-world datasets to demonstrate the effectiveness of our models on balancing the recommendation quality and user willingness.
