Modeling Bias Evolution in Fashion Recommender Systems: A System Dynamics Approach

Mahsa Goodarzi, M. Abdullah Canbaz · 2025

Read Modeling Bias Evolution in Fashion Recommender Systems: A System Dynamics Approach by Mahsa Goodarzi, M. Abdullah Canbaz (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-21728.txt.

Introduction

Mahsa Goodarzi and M. Abdullah Canbaz model how bias grows inside fashion recommender systems over time, rather than photographing it at one instant. In fashion shopping, skewed recommendations do more than annoy: they can entrench social stereotypes about bodies and styles, and each round of skewed suggestions trains the next round's data.

Using system dynamics modeling with experimental simulations, they separate the forces at play and trace how bias activates and reinforces itself. Inductive bias, the tilt built into the learning machinery itself, turns out to matter more than the bias users bring. Standard remedies, rebalancing data and regularizing algorithms, help only partly. The authors call for stronger strategies and wider system boundaries that include demographics and item diversity.

Catalog record

CREATORMahsa Goodarzi, M. Abdullah Canbaz
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-21728

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Modeling Bias Evolution in Fashion Recommender Systems: A System Dynamics Approach Mahsa Goodarzi, M. Abdullah Canbaz Abstract Bias in recommender systems not only distorts user experience but also perpetuates and amplifies existing societal stereotypes, particularly in sectors like fashion e-commerce. This study employs a dynamic modeling approach to scrutinize the mechanisms of bias activation and reinforcement within Fashion Recommender Systems (FRS). By leveraging system dynamics modeling and experimental simulations, we dissect the temporal evolution of bias and its multifaceted impacts on system performance. Our analysis reveals that inductive biases exert a more substantial influence on system outcomes than user biases, suggesting critical areas for intervention. We demonstrate that while current debiasing strategies, including data rebalancing and algorithmic regularization, are effective to an extent, they require further enhancement to comprehensively mitigate biases. This research underscores the necessity for advancing these strategies and extending system boundaries to incorporate broader contextual factors such as user demographics and item diversity, aiming to foster inclusivity and fairness in FRS. The findings advocate for a proactive approach in recommender system design to counteract bias propagation and ensure equitable user experiences. keywords Fashion Recommender Systems, Bias, Dynamic System Simulation 1 Introduction Recommender systems are a cornerstone of artificial intelligence applications, boasting a robust track record in the field. Hence, the recent shifts in research priorities have increasingly focused on enhancing algorithmic fairness and implementing bias mitigation strategies. While a substantial body of work has concentrated on optimizing accuracy and system performance, there remains a notable scarcity of studies specifically addressing the manifestations and impacts of bias within AI systems. Moreover, the exacerbation of these biases through feedback loops underscores the need for investigations into the dynamic, rather than static, nature of these systems in real-world scenarios Shi et al. 2024-03-31 ; Paparella et al. 2023-09-14 . In response to this gap, our paper explores the temporal dynamics of bias within Fashion Recommender Systems (FRS), employing system dynamics modeling and experimental simulations to understand and mitigate biases Malitesta et al. 2023-11-02 ; Mansoury et al. 2020-07-25 . We focus on FRS because their rapidly evolving trends, diverse user tastes, and strong visual/components expose biases more starkly than other domains. The analysis of human beauty standards and the historical changes in fashion products and their representations in the media reveal the deep-seated biases inherent in the fashion industry. These biases, which manifest as preferences for certain body shapes, skin tones, and facial features, often lead to negative body images and misrepresent diverse identities Ding 2023 . Indeed, the concept of bias in AI, as it pertains to fashion, is not limited to overt discriminatory behaviors but also includes preconceived limitations of AI in recognizing diverse forms of beauty, the intricate nature of cultural identities, and various data quality issues that affect outcome accuracy and objectivity Özer 2023 ; Pachouly 2024-01-31 ; Sai et al. 2024 . On the other hand, fashion recommendation systems(FRS) often decontextualize ‘fashion identity’ by prioritizing statistical objectives over the nuanced understanding of individual user identities. This reductionist approach can inadvertently reinforce biases, ignoring the complex interplay of gender, race, body shape, and cultural norms, which are crucial for enhancing user satisfaction and shaping brand perception Onitiu 2022 . In contrast, considering user feedback and socio-cultural contexts in the algorithmic modeling process results in more inclusive and representative recommendations Deldjoo et al. 2023-09-12 . In addition, the increase in diversity of recommendations and product exposure can enhance customer engagement and help mitigate inherent social biases. However, there exists a delicate balance between maintaining system accuracy and achieving fairness. Integrating personality-based data into FRS can potentially enhance user satisfaction, albeit at the potential cost of reduced precision Paryudi et al. 2022 . Furthermore, feedback loops in FRS can amplify existing biases, such as popularity and presentation biases, through the recurrent use of biased data and interactions. By incorporating multiple types of interactions—ranging from individual item selections to entire outfit choices—and contextualizing these within user-specific data frameworks, FRS can begin to effectively counteract these biases Celikik et al. 2022-09-18 ; Werneck et al. 2022 ; Wu et al. ; Elsayed et al. 2022-05-05 . Please note that we assume that any debiasing intervention (e.g., re‐balancing, regularization) yields a net improvement in recommendation quality. This assumption guides our feedback‐loop analysis; its limitations are noted in Section 5. In this paper, we aim to make a significant contribution to the literature by addressing the temporal dynamics of biases within Fashion Recommender Systems (FRS). Through system dynamics modeling and experimental simulations, we have developed a comprehensive conceptual framework that identifies and mitigates the reinforcing effects of biases using a multidimensional feedback approach. By integrating user, item, and interaction data, our dynamic framework systematically explores and addresses potential origins of bias while adapting to evolving user behaviors and shifting market trends. Please note that we specifically chose FRS due to their unique blend of rapidly changing trends, diverse user preferences, and complex item-user interactions. However, our model is also applicable to other domains, such as music, movie, and e-commerce recommender systems, where similar dynamics impact recommendation quality and fairness. Several recent works address dynamic biases in recommender systems. For instance, Zhang et al. 2020-03 proposes a probabilistic framework to disentangle popularity and position biases through counterfactual evaluation in a dynamic setting. Similarly, Dean et al. 2024 develops a graph‐based method that tracks bias evolution over time by continuously updating user–item interactions. However, these approaches focus primarily on localized bias estimation rather than capturing the full feedback‐loop dynamics across data, user behavior, and algorithm updates. In contrast, our system‐dynamics model explicitly simulates the nonlinear interactions among data‐ and design‐induced biases, user interactions, and recommendation performance within Fashion Recommender Systems. 2 Key Concepts: Bias and Recommendation Quality in Fashion Recommender Systems To effectively address bias and enhance recommendation quality in Fashion Recommender Systems (FRS), it is crucial to first clarify the key concepts that underpin this study. Bias plays a central role in shaping the performance and overall quality of recommender systems, influencing various nodes within these systems. In this section, we define and contextualize the different types of biases relevant to FRS, considering both explicit and unintentional forms. By clearly distinguishing these biases and exploring their implications on recommendation quality, we lay the foundation for understanding the dynamics of bias within FRS and justify the modeling choices and mitigation strategies discussed later in the paper. Bias in Recommender Systems: Bias in recommender systems refers to systematic and unfair tendencies that influence the recommendations presented to users. In the context of Fashion Recommender Systems (FRS), biases can stem from various sources, each with unique implications. Position bias, for example, occurs when items displayed in prominent positions (e.g., at the top of a recommendation list) receive disproportionately more attention than those listed further down Collins et al. 2018 . This bias is particularly impactful in FRS, where visual placement heavily influences user interactions and purchasing decisions. Another common issue is popularity bias, where algorithms tend to favor items that are already popular, further boosting their visibility while sidelining niche or diverse styles Kowald and Lacic 2022 . This homogenization effect limits exposure to underrepresented products, reinforcing trends that may not align with the varied tastes

Read the complete text