[Paper Review] An Algorithmic Framework to Control Bias in Bandit-based Personalization

L. Elisa Celis, Sayash Kapoor+2 more|arXiv (Cornell University)|Feb 23, 2018

Advanced Bandit Algorithms Research12 references15 citations

TL;DR

This paper proposes a novel algorithmic framework to control bias in bandit-based personalization by enforcing fairness constraints on content group exposure through a hierarchical group structure. It introduces two scalable bandit algorithms—Fair-EPS and Fair-OFUL—that achieve low regret while maintaining fairness, with experiments showing only a 20x% revenue loss to achieve a risk difference of less than 1−x for x<1/2.

ABSTRACT

Personalization is pervasive in the online space as it leads to higher efficiency and revenue by allowing the most relevant content to be served to each user. However, recent studies suggest that personalization methods can propagate societal or systemic biases and polarize opinions; this has led to calls for regulatory mechanisms and algorithms to combat bias and inequality. Algorithmically, bandit optimization has enjoyed great success in learning user preferences and personalizing content or feeds accordingly. We propose an algorithmic framework that allows for the possibility to control bias or discrimination in such bandit-based personalization. Our model allows for the specification of general fairness constraints on the sensitive types of the content that can be displayed to a user. The challenge, however, is to come up with a scalable and low regret algorithm for the constrained optimization problem that arises. Our main technical contribution is a provably fast and low-regret algorithm for the fairness-constrained bandit optimization problem. Our proofs crucially leverage the special structure of our problem. Experiments on synthetic and real-world data sets show that our algorithmic framework can control bias with only a minor loss to revenue.

Motivation & Objective

• To address the problem of systemic bias and polarization in bandit-based personalization systems that arise from over-specialization in content selection.
• To develop a general, extensible framework that allows users to specify custom fairness constraints on content groups without fixing a single fairness metric.
• To design a scalable, low-regret algorithm for constrained bandit optimization that maintains high revenue while ensuring fairness in content exposure.
• To empirically evaluate the trade-off between fairness and revenue performance on synthetic and real-world datasets, including the YOW dataset.

Proposed method

• Formalizes online group fairness by imposing linear upper and lower bounds on the probability of selecting content from each group, ensuring diversity in content exposure.
• Models the content selection problem as a constrained optimization task where fairness constraints are encoded as bounds on group-level probabilities.
• Adapts existing unconstrained bandit algorithms (e.g., ε-greedy and OFUL) by integrating a hierarchical, bottom-up probability allocation procedure that respects group constraints.
• Uses a recursive, tree-based allocation strategy: first satisfy lower bounds at the leaf level (individual arms), then allocate remaining probability to the highest-reward arm while respecting upper bounds.
• Employs dynamic bound adjustments: if a parent group’s total probability exceeds the sum of its children’s upper bounds, the parent’s upper bound is reduced accordingly.
• Proves that the greedy allocation strategy is optimal by contradiction, showing that any alternative allocation could be improved by reallocating to the highest-reward arm without violating constraints.

Experimental results

Research questions

• RQ1 Can fairness constraints be formally integrated into bandit-based personalization without significantly degrading revenue or regret?
• RQ2 How can a general framework be designed to support diverse fairness metrics (e.g., risk difference, representation disparity) through customizable group constraints?
• RQ3 What is the trade-off between fairness and revenue in real-world and synthetic bandit settings under constrained optimization?
• RQ4 How do hierarchical group structures and reward distributions affect the performance and scalability of fairness-aware bandit algorithms?

Key findings

• The Fair-EPS and Fair-OFUL algorithms achieve low regret and converge quickly to the theoretical optimum in the constrained setting on both synthetic and real-world data.
• For a risk difference constraint of less than 1−x (where x<1/2), the algorithm incurs only a 20x% revenue loss, demonstrating a manageable trade-off between fairness and utility.
• Even under the tightest fairness constraints (e.g., risk difference = 0), the optimal constrained reward remains within a factor of 2 of the unconstrained maximum reward.
• The proposed greedy allocation algorithm runs in O(gk) time, where g is the height of the group hierarchy and k is the number of arms, making it scalable for web-scale applications.
• The framework successfully maintains fairness by preventing over-specialization—content exposure remains diversified across groups even as the system learns user preferences.
• Experiments show that the algorithmic framework can effectively control bias while preserving high user engagement, suggesting practical viability for real-world deployment in news feeds and ad systems.