Open Access

ARTICLE

Enhancing Personalized Fashion Recommendation by Integrating Large Language Models with Attribute Features

Ti-Lun Miao¹, Hsien-Tsung Chang^1,2,3,*

1 Department of Computer Science and Information Engineering, Chang Gung University, Taoyuan, Taiwan
2 Department of Artificial Intelligence, Chang Gung University, Taoyuan, Taiwan
3 Center for Artificial Intelligence in Medicine, Chang Gung Memorial Hospital at Linkou, Taoyuan, Taiwan

* Corresponding Author: Hsien-Tsung Chang. Email:

(This article belongs to the Special Issue: Advances in Natural Language Processing and Large-scale AI Models)

Computer Modeling in Engineering & Sciences 2026, 148(2), 31 https://doi.org/10.32604/cmes.2026.086762

Received 05 June 2026; Accepted 12 August 2026; Issue published 28 August 2026

Abstract

Personalized fashion recommendation requires models that can capture visual compatibility, textual semantics, structured attributes, and user-specific preferences. However, existing multimodal approaches often rely on static word embeddings and shallow text encoders, limiting their ability to represent nuanced fashion descriptions. This study proposes a multimodal recommendation framework enhanced by large language models (LLMs) that integrates visual features, contextual textual representations, and structured attribute features for personalized outfit matching. A Japanese pretrained BERT encoder is used to replace the conventional Word2Vec and convolutional neural network (CNN)-based text pipeline, while GPT-4o is employed to extract fine-grained fashion attributes from product metadata. In addition, Llama-3.3-70B-Instruct is used to estimate semantic similarity among attribute values, enabling attribute-aware compatibility modeling beyond exact matching. Experiments on the IQON3000 dataset, containing 216,791 top-bottom outfit combinations, show that the proposed model achieves an area under the receiver operating characteristic curve (AUC) of 0.8477, outperforming the original Personalized Outfit Recommendation Scheme with Attribute-wise Interpretability based on Bayesian Personalized Ranking (PAI-BPR) baseline. Ranking-based evaluation further demonstrates that the proposed model consistently outperforms PAI-BPR across different candidate-set sizes and places compatible items closer to the top of the recommendation list. These results demonstrate that integrating large language models with multimodal and structured attribute features can effectively improve the accuracy and personalization of fashion recommendation systems.

---

Keywords

---