Quantifying Popularity Bias in Recommender Systems

This paper addresses the critical issue of quantifying popularity bias in recommender systems (RecSys), emphasizing the importance of fair and safe commercial applications. Four metrics are proposed to measure popularity bias over time across different user groups, providing comprehensive insights into disparities in treatment. By extending existing metrics and introducing new ones, such as Within-group-Gini coefficient and Dynamic-ΔGAP, the study sheds light on biases emerging over time asymmetrically among user groups. Demonstrated on commonly used datasets, the proposed metrics offer valuable tools for assessing and tracking biases in RecSys, essential for ensuring fairness and inclusivity in recommendation systems.

Addressing Popularity Bias in Recommender Systems (RecSys):

• Emphasizes the need to quantify biases in RecSys for fair and safe commercial applications, highlighting the potential adverse effects on individuals and society.
• Proposes comprehensive metrics to measure popularity bias over time across diverse user groups, providing a nuanced understanding of disparities in treatment.

Extending Existing Metrics and Introducing New Ones:

• Extends commonly used metrics such as Gini coefficient and ΔGAP to dynamic settings and different user groups, ensuring a more comprehensive assessment of biases.
• Reveals biases that emerge over time asymmetrically among user groups, offering insights crucial for understanding and addressing fairness issues.

Demonstrated Effectiveness:

• Demonstrates the proposed metrics’ effectiveness using commonly used datasets, underscoring their relevance and applicability in real-world scenarios.

Ensuring Fairness and Inclusivity:

• Provides valuable tools for assessing and tracking biases in RecSys, essential for promoting fairness and inclusivity in recommendation systems.

arXiv link