Deconstructing the Filter Bubble: User Decision-Making and Recommender Systems

Abstract

We study a model of user decision-making in the context of recommender systems via numerical simulation. Our model provides an explanation for the findings of Nguyen, et. al (2014), where, in environments where recommender systems are typically deployed, users consume increasingly similar items over time even without recommendation. We find that recommendation alleviates these natural filter-bubble effects, but that it also leads to an increase in homogeneity across users, resulting in a trade-off between homogenizing across-user consumption and diversifying within-user consumption. Finally, we discuss how our model highlights the importance of collecting data on user beliefs and their evolution over time both to design better recommendations and to further understand their impact.

Simulation as Experiment: An Empirical Critique of Simulation Research on Recommender Systems

“…In Authors' addresses: Amy A. Winecoff, Princeton University, Princeton, New Jersey, USA, aw0934@princeton.edu; Matthew Sun, Princeton University, Princeton, New Jersey, USA, mdsun@princeton.edu; Eli Lucherini, Princeton University, Princeton, New Jersey, USA, elucherinin@cs.princeton.edu; Arvind Narayanan, Princeton University, Princeton, New Jersey, USA, arvindn@cs.princeton.edu. these cases, simulations are similar to empirical laboratory experiments, which eschew the complexities of the real world in order to characterize the theoretical causal relationship between variables that are often difficult to cleanly isolate in real-world settings [1]. Unlike the Bay model, the target system for many simulation models is an experiment.…”

“…Several recent studies [1,7,16,19] have used different simulation designs to study "filter bubbles" in RS, whereby users are exposed to an increasingly narrow range of content that matches their existing preferences, further homogenizing their interests [22]. The simulations in Chaney et al [7] and Mansoury et al [19] suggest that algorithmic feedback loops can over-homogenize user preferences [7] and exacerbate algorithmic bias towards popular items and away from the minority subpopulation's interest [19].…”

“…The simulations in Chaney et al [7] and Mansoury et al [19] suggest that algorithmic feedback loops can over-homogenize user preferences [7] and exacerbate algorithmic bias towards popular items and away from the minority subpopulation's interest [19]. In contrast, the simulations in Jiang et al [16] and Aridor et al [1] locate the cause of filter bubbles in the preferences and behavior of users and suggest that RS can either alter the rate at which users' interest homogenize [16] or prevent the homogenization of users' interest that would occur in the absence of recommendation [1]. Both Jiang et al [16] and Aridor et al [1] centrally incorporate some notion of user preference bias.…”

Recommendation System Simulations: A Discussion of Two Key Challenges

“…Some simulations assume that users are uncertain of their preferences. Aridor et al [1] model users as considering items sequentially, updating their beliefs as they go; under this paradigm, recommendations amount to adjusting the values which users rely on for decisions. Other work frames the utility of an item for a user as being comprised of two parts: known and unknown to the user; a user then acts based on a function of the recommended rank of an item and their known utility for that item [2].…”

“…Some work offers no alternatives to recommendations, focusing only on generating data with bias and the corresponding implications [18,19]. Others compare different recommendation systems such as random recommendations, recommendations based on overall popularity, matrix factorization, and recommendations under ideal or oracle conditions [1,2,9]. Along this vein, Geschke et al [6] compares recommending close content to distant content, but also includes social dynamics in their simulations.…”