[Paper Review] Exploring and Mitigating Gender Bias in Recommender Systems with Explicit Feedback

Shrikant Saxena, Shweta Jain|arXiv (Cornell University)|Dec 5, 2021

Recommender Systems and Techniques4 citations

TL;DR

This paper proposes a model to quantify and mitigate gender bias in recommender systems using explicit feedback, by debiasing user ratings through user-specific bias estimation and scaling. It demonstrates that debiased ratings produce significantly reduced bias (up to 40%) with minimal accuracy loss across multiple algorithms, including UserKNN, ItemKNN, ALS, and SVD, while remaining model-agnostic and theoretically grounded in unbiased estimation.

ABSTRACT

Recommender systems are indispensable because they influence our day-to-day behavior and decisions by giving us personalized suggestions. Services like Kindle, Youtube, and Netflix depend heavily on the performance of their recommender systems to ensure that their users have a good experience and to increase revenues. Despite their popularity, it has been shown that recommender systems reproduce and amplify the bias present in the real world. The resulting feedback creates a self-perpetuating loop that deteriorates the user experience and results in homogenizing recommendations over time. Further, biased recommendations can also reinforce stereotypes based on gender or ethnicity, thus reinforcing the filter bubbles that we live in. In this paper, we address the problem of gender bias in recommender systems with explicit feedback. We propose a model to quantify the gender bias present in book rating datasets and in the recommendations produced by the recommender systems. Our main contribution is to provide a principled approach to mitigate the bias being produced in the recommendations. We theoretically show that the proposed approach provides unbiased recommendations despite biased data. Through empirical evaluation on publicly available book rating datasets, we further show that the proposed model can significantly reduce bias without significant impact on accuracy. Our method is model agnostic and can be applied to any recommender system. To demonstrate the performance of our model, we present the results on four recommender algorithms, two from the K-nearest neighbors family, UserKNN and ItemKNN, and the other two from the matrix factorization family, Alternating least square and Singular value decomposition.

Motivation & Objective

• To quantify gender bias in book rating datasets and in recommendations derived from explicit feedback.
• To address the self-perpetuating cycle of bias in recommender systems that arises from biased user ratings.
• To develop a model-agnostic method that produces unbiased recommendations even when trained on biased data.
• To ensure fairness in recommendations by debiasing user ratings before feeding them into any recommender algorithm.
• To evaluate the method’s effectiveness across diverse algorithms while preserving recommendation accuracy.

Proposed method

• The model estimates each user’s gender-based bias by comparing their ratings of a book to the scaled average ratings given by peers to the same book.
• It applies a debiasing transformation to individual ratings using a user-specific bias score, producing a debiased rating that is an unbiased estimator of the true preference.
• The debiased ratings are then used as input to any standard recommender algorithm, ensuring fairness in downstream recommendations.
• The method is hybrid, combining pre-processing (rating debiasing) and post-processing (fair recommendation generation) phases.
• It uses statistical significance testing (z-test) to validate bias reduction across multiple datasets and algorithms.
• The approach is model-agnostic and can be applied to any collaborative filtering or matrix factorization-based recommender system.

Experimental results

Research questions

• RQ1 How can gender bias in explicit feedback from users be quantified in book recommender systems?
• RQ2 To what extent can user-specific bias be estimated and corrected to produce fairer recommendations?
• RQ3 Does debiasing user ratings lead to a significant reduction in recommendation bias without compromising accuracy?
• RQ4 How does the performance of the debiasing model vary across different recommender algorithms such as UserKNN, ItemKNN, ALS, and SVD?
• RQ5 What is the impact of dataset size and user/item diversity on the accuracy of bias estimation and mitigation?

Key findings

• The proposed method reduced gender bias by up to 40% in recommendations, with the most significant reduction observed in UserKNN-based systems.
• Bias reduction was more pronounced in the AZ dataset due to its larger user and item base, enabling more accurate bias score estimation.
• The model achieved a 10% accuracy loss with UserKNN while reducing bias by nearly 40%, indicating a favorable trade-off between fairness and performance.
• Statistical significance tests confirmed that the reduction in bias was highly significant (p < 10^-5) across all algorithms and datasets.
• The method performed better with UserKNN than ItemKNN, as it aligns more naturally with the user-based similarity mechanism used in UserKNN.
• Accuracy loss was higher on the smaller BX dataset, confirming that larger datasets improve the reliability of bias estimation and mitigation.