Algorithmic bias and fairness in recommendation systems

Loading...
Thumbnail Image

Date

Journal Title

Journal ISSN

Volume Title

Publisher

Setif 1 University - Ferhat ABBAS , Faculty of Sciences

Abstract

Recommender systems (RSs) have become essential components of modern digital platforms, en-abling personalized delivery of content, products, and services across domains such as e-commerce, streaming, social media, and online news. By leveraging large-scale user interaction data, they substantially influence information discovery and user engagement. However, RSs often inherit and amplify biases embedded in historical data, raising significant fairness concerns. These biases arise along two principal dimensions. On the user side, recommendations may become correlated with sensitive demographic attributes, such as gender, leading to stereotypical or discriminatory out-comes. On the item side, popularity bias reinforces the overexposure of highly popular items while systematically limiting the visibility of long-tail content, thereby reducing catalog diversity.As inherently multi-stakeholder systems, recommender systems must balance the interests of users seeking relevant recommendations and item providers requiring equitable exposure. Consequently, fairness is fundamentally a two-sided problem. Existing approaches typically address either user-side or item-side fairness in isolation or apply fairness constraints only to recommendation outputs, leaving biased latent representations unchanged. Such representations continue to encode sensitive information, allowing bias to re-emerge. This dissertation argues that effective and durable two-sided fairness must be achieved at the representation level by learning unbiased user and item embeddings before recommendation generation. This research proposes a representation-centric framework for two-sided fairness that directly debi-ases latent user and item representations. The proposed approach simultaneously mitigates user-side demographic bias and item-side popularity bias while preserving recommendation quality. Extensive experiments on multiple benchmark datasets demonstrate significant improvements in both fairness dimensions without compromising recommendation accuracy. The proposed approach consistently outperforms state-of-the-art fairness-aware recommender systems across a broad range of fairness metrics, highlighting the effectiveness of representation-level debiasing for developing fair, accurate, and balanced recommender systems.

Description

Citation

Endorsement

Review

Supplemented By

Referenced By