面向国潮文化偏好的多模态电商推荐技术综述

A Review of Multimodal E-Commerce Recommendation Technologies for Domestic Trend Cultural Preferences

摘要: 随着国潮品牌与文化消费的持续发展，电商平台中的商品推荐面临由功能与价格导向向文化符号、审美风格与情感认同导向转变的趋势。国潮商品通常同时包含视觉风格、文本语义与文化属性等多源信息，单一模态或传统协同过滤方法难以有效刻画其复杂的偏好特征。近年来，基于深度学习的多模态推荐技术在电商领域取得了广泛应用，为融合商品图像、文本描述、用户评论及结构化属性信息提供了新的研究思路，但其在国潮文化偏好建模中的系统性总结仍相对不足。本文围绕面向国潮文化偏好的多模态电商推荐技术展开综述。首先分析国潮消费场景下用户偏好与商品特征的多模态特性，梳理相关数据形态与常用公开数据资源；随后从多模态表征学习、跨模态融合与对齐、图结构建模与知识增强、模型优化与训练策略四个维度，对近年来的代表性研究工作进行分类总结，并探讨各方法在国潮场景下的适配性。在此基础上，本文探讨了大语言模型驱动的生成式推荐范式，分析其在深层文化知识推理与可信解释生成方面的破局作用。最后，本文总结了当前研究在文化语义建模、模态对齐、偏好动态建模与推荐解释等方面面临的主要挑战，并对多模态大模型驱动的电商推荐、人机协同交互机制以及文化偏好可控建模等未来研究方向进行了展望。本文可为国潮商品推荐系统的研究与实践提供参考。

Abstract: With the continued growth of domestic trend brands and cultural consumption, product recommendations on e-commerce platforms are shifting from a focus on functionality and price to an emphasis on cultural symbols, aesthetic styles, and emotional resonance. Domestic trend products typically involve multi-source information, including visual style, textual semantics, and cultural attributes. As a result, single-modality methods and traditional collaborative filtering approaches often struggle to capture such complex preference patterns effectively. In recent years, deep learning-based multimodal recommendation techniques have been widely adopted in e-commerce, providing new avenues for integrating product images, textual descriptions, user reviews, and structured attribute information. However, systematic reviews of their applications to modeling preferences in domestic trend culture remain limited. This paper presents an overview of multimodal recommendation technologies for e-commerce that are tailored to domestic trend cultural preferences. First, we examine the multimodal characteristics of user preferences and product attributes in domestic trend consumption scenarios, and we identify relevant data modalities, formats, and commonly used public datasets. We then categorize and summarize representative studies from recent years across four dimensions: multimodal representation learning, cross-modal fusion and alignment, graph structure modeling and knowledge enhancement, and model optimization with training strategies, while discussing the suitability of these methods for domestic trend scenarios. Furthermore, this paper highlights the LLM-driven generative recommendation paradigm, analyzing its breakthrough role in deep cultural knowledge reasoning and the generation of trustworthy explanations. Finally, we summarize key challenges in cultural semantic modeling, modality alignment, preference dynamics modeling, and recommendation explainability. We also outline promising future directions, such as multimodal large-model-driven e-commerce recommendation, human-machine collaborative interaction mechanisms, and controllable modeling of cultural preferences. This paper aims to serve as a reference for both research and practice in recommendation systems for domestic trend products.