In a study published in the journal Fashion and Textiles, researchers demonstrated that large language models can generate content-based explanations that significantly enhance user satisfaction and trust in fashion recommendation systems. The research, conducted by Haein Yeo, Taehyung Noh, and Kyungsik Han, addresses a common limitation in e-commerce: the reliance on user reviews for explaining why specific items are recommended. By leveraging visual attributes and metadata extracted from product images, the team developed a framework that produces tailored textual explanations without requiring existing review data. This approach aims to make AI-driven fashion recommendations more transparent and understandable for consumers who may not have established preferences or extensive interaction histories with a platform.

The study was motivated by the observation that users often experience uncertainty when interacting with recommender systems, particularly when the reasoning behind a recommendation is opaque. In the fashion domain, where personal style and subjective taste play a central role, this uncertainty can hinder decision-making. Previous approaches to explainability often relied on review-based methods, which are limited by the availability and quality of user-generated content. Products with few or no reviews, or those with low-quality feedback, could not benefit from these explanations. Furthermore, visual elements, which are critical to fashion purchase decisions, were frequently neglected in traditional explanation generation methods.

To overcome these challenges, the researchers developed a web-based interface called My Own Style, or MOS, to conduct a controlled user study. The system utilized deep learning models to extract three distinct types of information from fashion product images: visual embeddings, fashion attributes and categories, and dominant colors. The visual information was captured using a Vision Transformer-Base model, which processes the entire image to create a 768-dimensional vector representing the product’s overall shape and style. Simultaneously, an Attribute Mask R-CNN model detected specific fashion attributes and categories, utilizing a predefined ontology of 46 categories and 294 attributes derived from the Fashionpedia dataset. Color information was extracted from the largest detected bounding box and embedded as a 3-dimensional RGB vector.

The extracted data was then fed into a Large Language Model, specifically GPT-3.5-turbo, to generate natural language explanations. The researchers designed two levels of explanation detail: a simple level that used only category and color information, and a detailed level that incorporated specific fashion attributes. These content-based explanations were compared against a basic, conventional explanation type typically found in e-commerce interfaces, which often provides little to no rationale for the recommendation. The prompts were carefully engineered to ensure the LLM produced human-like, natural descriptions that avoided raw data formats, such as RGB values, and instead used descriptive language understandable to general users.

The recommendation engine employed two distinct algorithms to test how explanation detail interacts with recommendation strategy. The first was a similarity-based algorithm, which suggested products that closely matched the visual characteristics of the user-selected item. The second was a diversity-based algorithm, which offered items with moderate similarity but sufficiently diverse styles to encourage exploration. This diversity approach was designed to mitigate the filter bubble effect, where users are confined to a narrow set of options that merely mirror their past preferences. By using a content-based filtering approach, the system could generate recommendations and explanations for all products, regardless of their review history.

The user study involved 211 participants recruited from Amazon Mechanical Turk, with criteria ensuring they had experience purchasing clothing online. Each participant completed 18 recommendation instances, experiencing all six experimental conditions derived from the two algorithms and three explanation levels. The study used a within-subjects design, allowing researchers to measure how individual perceptions changed across different combinations of algorithms and explanation types. Participants rated their satisfaction with the recommendations, their trust in the system, the transparency of the explanations, and their satisfaction with the explanations themselves, using 7-point Likert scales.

The results indicated that content-based explanations significantly improved user perceptions compared to basic explanations. Both simple and detailed explanations led to higher scores in recommendation satisfaction, trust, and explanation transparency. However, the effectiveness of the explanations varied depending on the recommendation algorithm. For the similarity-based algorithm, users showed significant improvements in satisfaction and trust when moving from basic to simple or detailed explanations. In contrast, the diversity-based algorithm showed the highest satisfaction and trust specifically when paired with detailed explanations. This suggests that when recommendations are less intuitive or more exploratory, users require more comprehensive information to understand the rationale and feel confident in the results.

The interaction between algorithm type and explanation level was statistically significant for recommendation satisfaction and trust. The detailed explanations helped users appreciate the diversity in the recommendations, reducing the uncertainty associated with unexpected items. This finding highlights the importance of tailoring the level of explanation detail to the specific characteristics of the recommendation algorithm to maximize user engagement and trust.

The study also found that explanation transparency was primarily driven by the level of explanation detail rather than the algorithm type. Both simple and detailed content-based explanations were perceived as significantly more transparent than basic explanations, with no significant difference between the simple and detailed levels for this specific metric. This indicates that providing any form of feature-based rationale helps users understand how the recommendation was derived, but the depth of that rationale is less critical for transparency than it is for trust and satisfaction in the context of diverse recommendations.

Limitations of the study include the use of single-item scales to measure complex psychological constructs like trust and satisfaction, which may not capture the full nuance of user perception as effectively as multi-item scales. Additionally, the participant pool from Amazon Mechanical Turk may not be fully representative of the general online shopping population, potentially influencing the demographic characteristics of the results. The dataset predominantly consisted of products from fast fashion brands, which may limit the generalizability of the findings to luxury or sustainable fashion segments where consumer decision-making processes differ. The researchers also noted that the LLM occasionally produced outputs that replicated prompt instructions or included raw data, requiring careful prompt engineering and review to ensure quality.

Despite these limitations, the study provides valuable insights for the design of explainable AI systems in e-commerce. It demonstrates that LLMs can be effectively used to generate content-based explanations that complement or replace review-based methods, particularly for products with limited user feedback. The findings suggest that e-commerce platforms can improve user experience by dynamically adjusting the level of explanation detail based on the recommendation algorithm being used. This approach can foster a more transparent and trustworthy shopping environment, empowering users to make informed decisions and explore their personal style more effectively.

The research contributes to the fields of human-computer interaction, artificial intelligence, and fashion studies by offering a methodology for generating fashion-specific LLM prompts. It highlights the potential for AI to support meaningful choices for consumers online by providing clear, context-aware explanations. As AI-powered services become more prevalent, the ability to explain the reasoning behind recommendations will be crucial for maintaining user trust and satisfaction. This study offers a practical framework for implementing such systems, emphasizing the need to balance personalization with transparency and to consider the specific needs of both customers and vendors in the design of recommendation explanations.

Subject of Research: Llm-generated content-based explanations for user experience in fashion recommender systems

Article Title: Llm-generated content-based explanations for user experience in fashion recommender systems

Article References: Yeo, H., Noh, T., & Han, K. (2026). Llm-generated content-based explanations for user experience in fashion recommender systems. Fashion and Textiles, 13(1), Article 16. https://doi.org/10.1186/s40691-026-00464-z

Image Credits: AI Generated

DOI: 10.1186/s40691-026-00464-z

Keywords: Llm-generated, content-based, explanations, user, experience, fashion, recommender, systems, scientific research