Fairness-in-the-Workflow: How Machine Learning Practitioners at Big Tech Companies Approach Fairness in Recommender Systems

Paper Title

Fairness-in-the-Workflow: How Machine Learning Practitioners at Big Tech Companies Approach Fairness in Recommender Systems

Publication Info

• Topic area: Fairness practices in recommender systems (RS) within large technology companies.
• Keywords: Fairness, recommender systems, machine learning, bias, multi-stakeholder, workflows, organizational challenges, fairness metrics, cross-team collaboration, HCI.

Background and Problem

• Problem / challenge: Despite extensive academic research on fairness in machine learning, translating these theories into practical applications for recommender systems (RS) remains challenging. RS practitioners face difficulties in defining fairness, balancing multi-stakeholder interests, and addressing fairness in dynamic environments.
• Significance: RS are widely deployed in high-stakes domains like e-commerce and social media, where biases can lead to significant societal impacts, such as monopolistic practices or ideological polarization.
• Motivation and related work: Prior work has focused on fairness in general-purpose ML systems, classification models, and NLP, but RS-specific challenges—such as multi-stakeholder dynamics and feedback loops—remain underexplored. This study builds on prior research by focusing on how RS practitioners at large tech companies incorporate fairness into their workflows.

Solution

• Proposed approach: A semi-structured interview study (N=11) with RS practitioners at large tech companies to map their workflows and identify challenges and strategies for incorporating fairness.
• Novelty:
  1. A detailed map of RS practitioner workflows, highlighting fairness considerations both internally and in collaboration with legal, data, and fairness teams.
  2. Identification of key technical and organizational challenges specific to fairness in RS workflows.
  3. Actionable recommendations for practitioners and HCI researchers to improve fairness integration in RS.
• Procedure and key techniques:
  • Conducted 11 semi-structured interviews with RS practitioners from seven large tech companies.
  • Analyzed workflows and challenges using inductive thematic analysis.
  • Mapped fairness considerations across three phases: prototyping, internal, and collaborative.

Results

• Concrete findings:
  • Practitioners spend less than 10% of their time on fairness tasks.
  • Fairness metrics are often intuition-based and developed ad hoc, with practitioners creating over five new metrics per project.
  • Practitioners struggle to define fairness, balance multi-stakeholder interests, and manage fairness in dynamic RS environments.
  • Cross-team communication is hindered by a lack of shared terminology and late-stage involvement of fairness teams.
• Advantage over baselines: This study provides a unique focus on RS workflows and fairness challenges across multiple large tech companies, unlike prior studies that focus on single organizations or general ML systems.
• Experiments / evaluation:
  • Semi-structured interviews with 11 practitioners averaging 5.36 years of experience.
  • Workflow mapping and thematic analysis to identify technical and organizational challenges.
• Limitations and future work:
  • Limited sample size (N=11) from seven companies; findings may not generalize to smaller-scale RS.
  • Focused only on technical practitioners, excluding perspectives from legal, data, and fairness teams.
  • Future work should explore cross-team alignment and develop tools to support fairness in RS workflows.

Summary

This study investigates how RS practitioners at large tech companies incorporate fairness into their workflows. Through semi-structured interviews, the authors identify technical challenges, such as defining fairness and managing multi-stakeholder interests, and organizational challenges, including limited time and poor cross-team communication. The study proposes actionable recommendations, such as improving documentation, integrating fairness early in workflows, and developing a shared "fairness lingua franca." These findings provide valuable insights for practitioners and HCI researchers aiming to improve fairness in RS, particularly in large-scale, high-stakes applications.