Quantifying Attribute Association Bias in Latent Factor Recommendation Models

Our evaluation framework for analyzing attribute association bias contributes to the research area of fairness and bias in recommendations. Additionally, we build upon seminal work in addressing gender attribute association bias in natural language processing. In this section, we provide an overview of relevant key findings and research.

2.1 Fairness and Bias in Recommendations

Understanding how to define and evaluate fairness and bias of recommendation systems has quickly grown into a seminal area of information retrieval research. Various types of bias related to recommendation systems have been defined and studied in academic and industry settings. Researchers often target studying bias relating to harms of allocation, unequal distribution of exposure, or attention of recommendations within the system [23]. Allocative, or distributional, harms have been studied by evaluating and mitigating biases such as popularity, exposure, ranking (or pairwise), and gender bias [2–4, 19, 25, 29, 40]. A recent literature review by Ekstrand et al. [23] notes that representational harms can also be studied in recommendation systems but focuses on representation in terms of the provider and how stakeholders view their distribution within the system, not their numerical representation as vector outputs of a recommendation system.

In this paper, we explore representation harm in terms of association bias, more commonly studied in NLP, to understand how stereotypes can become encoded into the latent embedding space [13, 23]. Previously, in NLP research, this bias was referred to as gender association bias [13]. Our proposed framework and methodologies build upon previous NLP methods for gender association bias to analyze association bias agnostic of the type of bias in recommendation settings. Our modifications account for distinct differences between NLP and LFR models and generalize beyond gender when evaluating association bias.

Our proposed methodologies are designed to help uncover implicit or explicit attribute association bias that can affect latent factor recommendation models and may unknowingly amplify stereotypes or contribute to downstream allocation harms. Our work evaluating attribute association bias presents a new direction in exploring bias in recommendations by addressing the current gap in quantifying representation bias in the vector representations leveraged and outputted by a recommendation system. In addition to studying a form of bias not often addressed in recommendation system research, we also present metrics that address concepts not currently captured by recommendation fairness research. Current metrics target distributional harms resulting from the biases mentioned above by measuring the equity of recommendation results through accuracy based, pairwise based, or causal based methods [24, 57]. Instead of focusing on distributional outcomes, our techniques focus on increasing the transparency of entity relationships embedded into the latent space by evaluating entity bias associations. These relationships can be studied regarding individual and group fairness, where group fairness focuses on providing similar outcomes across groups of entities [22], whereas individual fairness specifies that similar entities should receive similar outcomes [22]. Our framework focuses on evaluating the group fairness of attribute association bias to understand if one group of entities experiences more stereotyped encodings than another.

2.1.1 Gender Bias. Our case study (§5) presents an offline evaluation of attribute association bias relating to user gender stereotypes, specifically how recommendations of specific pieces or types of content may vary according to a user’s gender. Beyond the scope of this work, offline evaluation of user gender bias in recommendations has been largely unexplored due to the difficulty in obtaining the user’s gender for analysis. User gender bias evaluation is often only available to industry practitioners, given their ability to collect and analyze user attribute data. However, results are only sometimes shared due to the sensitive nature of this work. We hope that by sharing our findings concerning user gender bias, we increase transparency within this space and encourage other industry researchers and practitioners to follow suit, thus allowing a greater ability for industry and academia to collaborate, address, and mitigate these kinds of challenges.

Given this, past research on evaluating gender bias in recommendations has more frequently focused on allocation harms occurring from provider or creator gender bias in the recommendation system. For example, Ekstrand et al. [25] explored the effects of author gender on book recommendations by exploring item recommendation distributions according to author gender, showing that recommendation algorithms could potentially recommend specific authors in a gender-biased fashion. Shakespeare et al. [50] evaluated the extent to which artist gender bias is present within music recommendations. Ferraro et al. [28] conducted experiments studying how collaborative filtering performs when specifically accounting for an artist’s gender in music recommendation. Our work differentiates from past research by exploring representational harms of user gender resulting from LFR models. Even though quantifying user gender bias in media recommendation systems remains rare, researchers have produced work leveraging qualitative user studies to evaluate gender differences in media preference. For example, Millar [42] conducted a user study of young female and male adults to evaluate "gender differences in artist preferences". Berkers [10] conducted an offline analysis of gender preferences in the Last.FM dataset, showcasing how data can capture gendered listening patterns in music.

2.2 Quantifying Gender Associations in Natural Language Processing Representations Our proposed evaluation framework for measuring entity bias associations in latent recommendation model embeddings is inspired by natural language processing (NLP) methods that attempt to measure binary gender bias in word embeddings — e.g., associations between gender-neutral words (like “scientist” or “nurse”) and words indicative of a specific gender (like “man” or “woman”). Past work has identified gender biases in pretrained static word embeddings [13, 15, inter alia], contextual word embeddings from large language models [8, 11, 53, 63], and embeddings of larger linguistic units like sentences [34, 38]. Because pretrained word & sentence embeddings are widely used as input for many NLP models, there is potential for biases in embeddings to be propagated or amplified in downstream text classification and generation tasks [46, 49, 63]. In the recommendation system setting, we consider the analogous possibility of sensitive feature associations in latent entity embeddings due to the possibility of propagation in downstream models [14]. However, unlike analysis of gender in relation to words, recommendation entities do not necessarily have natural contrastive pairings by attribute. We address this gap in §4.

Authors:

This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license.