Center for Social Information Sciences (CSIS) Seminar

Abstract: Recommendation systems are pervasive in the digital economy. An important assumption in many deployed systems is that user consumption reflects user preferences in a static sense: users consume the content they like with no other considerations in mind. However, as we document in a large-scale online survey, user behavior departs from this model in a crucial way: users choose content strategically, with the goal of influencing their future recommendations. We model the user behavior as a two-stage noisy signalling Stackelberg game between the recommendation system and users: in a finite-length first stage, the recommendation system implements a preliminary recommendation policy, to which users respond by strategically consuming content to change the breakdown of content types recommended to them in the future. Based on the user's preferences as learned in this first stage, in the second stage, the platform commits to a recommendation policy by which it will recommend content to the users. We show that users' strategic behavior can affect the user experience: at equilibrium in this game, differences in users' preferences become accentuated as they strategically consume content further outside the mainstream. This effect is particularly strong among users from (statistical) minorities, who must specifically avoid signaling interest in mainstream content in order to ensure the algorithm will show them content related to their minoritized identities. This puts minority users at an unfair disadvantage, where they cannot access mainstream content without the algorithm suppressing the content types that only they (as the minority) enjoy.

We next propose three interventions that improve the recommendation quality (both on average and for minority groups) that account for strategic consumption:(1) Adopting a recommendation system policy that uses preferences from a prior, (2) Communicating to users that universally liked ("mainstream") content will not be used as the basis of recommendation, and (3) Serving content that is personalized-enough yet expected to be liked in the beginning. Finally, we describe a methodology to inform applied theory modeling in incentive-aware learning settings with survey results.

Based on joint work with Andreas Haupt (MIT) and Dylan Hadfield-Menell (MIT).