Abstract
While popularity bias is recognized to play a role in recommender (and other ranking-based) systems, detailed analyses of its impact on user welfare have largely been lacking. We propose a general mechanism by which item popularity, item quality, and position bias can impact user choice, and how it can negatively impact the collective user utility of various recommender policies. Formulating the problem as a non-stationary contextual bandit, we highlight the importance of exploration, not to eliminate popularity bias, but to mitigate its negative effects. First, naive popularity-biased recommenders are shown to induce linear regret by conflating item quality and popularity. More generally, we show that, even in linear settings, identifiability of item quality may not be possible due to the confounding effects of popularity bias. However, under sufficient variability assumptions, we develop an efficient UCB-style algorithm and prove efficient regret guarantees. We complement our analysis with several simulation studies.
Introduction
In the world of recommender systems, a "rich-get-richer" phenomenon often dominates: items gain popularity not necessarily due to superior quality, but because they are already popular. This self-amplifying feedback loop, known as popularity bias, can lead algorithms to promote low-quality content simply because it was trending early on. This research explores how these dynamics impact user welfare and proposes a method to disentangle quality from popularity.
The Confounding of Choice
User choice in a ranking system is typically influenced by three main factors:
- Item Quality: The intrinsic merit of the item.
- Position Bias: The tendency to click higher-ranked items regardless of merit.
- Popularity Bias: The tendency to choose items that are already widely consumed.
The core challenge is that these factors are confounded. When a user selects a popular item at the top of a list, the system cannot easily determine if the choice was driven by quality or bias. This creates a non-stationary environment where the act of recommendation alters future popularity, potentially trapping the system in a suboptimal state.
The "Unlucky Draw" and Identifiability
A key insight of the paper is the "Impossibility Theorem," which states that without specific assumptions, a ranker cannot distinguish between a high-quality item and a low-quality item that received an early popularity boost. This leads to the "Unlucky Draw" scenario, where a mediocre item dominates the rankings indefinitely because it was selected early on.
The Quality-Popularity (QP) Ranker
To address these challenges, the authors introduce the Quality-Popularity (QP) Ranker. This algorithm is modeled as a non-stationary contextual bandit that uses an Upper Confidence Bound (UCB) approach.
- Decoupling Biases: The QP Ranker explicitly maintains separate estimates for quality and popularity parameters.
- Strategic Exploration: It leverages a "diversity" assumption—the idea that different user contexts allow items to be observed under varying conditions, helping the algorithm eventually "decouple" the biases.
- Saturation-Awareness: The ranker continues to explore until item popularity effects are sufficiently understood (saturated), preventing premature convergence on mediocre items.
Key Results
Simulation studies demonstrate that the QP Ranker achieves sublinear regret, meaning it eventually learns to maximize user satisfaction over time. In contrast, "greedy" or "popularity-oblivious" rankers often suffer from linear regret, failing to ever identify the best items due to the feedback loop.
The QP Ranker significantly outperforms baselines by successfully identifying and promoting the highest-quality items, even when they start with low popularity. This ensures that long-term social welfare is maximized rather than just following temporary trends.
Conclusion
This work provides a rigorous foundation for understanding and mitigating popularity bias. By treating recommendation as a non-stationary bandit problem and employing strategic exploration, we can build systems that prioritize true item quality over temporary hype. Future work will explore non-linear interactions between quality and popularity and the simultaneous learning of position bias.
Summary updated