ChatPreference: a large language model for customer preference understanding in the apparel industry

In the era of mass personalisation, understanding customer preferences (CPs) stands as a core requirement for product conceptual design. Products that fail to meet CPs are likely to lose competitiveness, particularly in the apparel industry, where clothing is a necessity and highly individualised. Traditional approaches rely on surveys, post-purchase reviews, or behavioural logs, which are labour-intensive and constrain the natural ways customers express their preferences. Leveraging the semantic understanding capability of large language models, this study develops ChatPreference through a three-stage training framework that enables CP analysis and tailored recommendations. Using men’s dress shirts as a representative category, the framework is implemented in three stages. First, basic training, a domain-specific knowledge corpus is constructed and used for knowledge pretraining to build a foundational understanding of shirt design attributes and usage scenarios. Second, professional training, a tailored Query & Answer dataset is constructed and augmented for supervised fine-tuning, enabling the model to interpret diverse CP expressions and generate personalised recommendations. Third, specialised training, model outputs are further refined through a Consistency-regularized Direct Preference Optimization (Core-DPO) method that considers labelling noise, thereby achieving closer alignment with human expectations. Lastly, benchmark evaluations demonstrate that ChatPreference outperforms state-of-the-art models in CP understanding and recommendation.