融合扩散模型的多模态图对比学习推荐

PDF下载HTML

Title:

Multimodal graph contrastive learning recommendation based on fusion diffusion models

作者:

沈学利, 于佳灏

辽宁工程技术大学 软件学院, 辽宁 葫芦岛 125105

Author(s):

SHEN Xueli, YU Jiahao

School of Software, Liaoning Technical University, Huludao 125105, China

关键词:

推荐系统; 扩散模型; 多模态; 图对比学习; 自监督学习; 多任务学习; 数据稀疏; 用户兴趣

Keywords:

recommender system; diffusion model; multimodal; graph contrastive learning; self-supervised learning; multitask learning; data sparsity; user interest

分类号:

TP311

DOI:

10.11992/tis.202507015

摘要:

针对多模态数据中固有噪声干扰、数据稀疏性以及跨模态语义表示差异导致的用户兴趣偏好建模不准确的问题，提出了一种融合扩散机制的图对比学习推荐方法(diffusion-enhanced exploration for multimodal graph contrastive recommendation，DEMGR)。借助扩散模型生成结构增强的对比视图，有效抑制自监督学习中各个模态噪声的干扰；构建融合用户与物品的同构图结构，增强图的连通性，缓解稀疏性对图表示学习的影响；设计一个联合优化框架，融合多任务学习与扩散增强机制，协同优化推荐目标与增强信号，学习具有语义一致性与泛化能力的多模态表示。在Baby、Sports和Clothing 共3个真实数据集上的实证研究表明，DEMGR在Recall和NDCG(normalized discounted cumulative gain)等指标上均显著优于现有主流模型，最大提升分别达5.45%和5.01%，充分验证了该方法在多模态推荐中的有效性。本文分析结果可为多模态推荐系统优化提供参考。

Abstract:

To address the inaccurate modeling of user interest preferences caused by noise interference, data sparsity, and cross-modal semantic discrepancies in multimodal data, this paper proposes a diffusion-enhanced exploration framework for multimodal graph contrastive recommendation(DEMGR). First, a diffusion model is employed to generate structurally enhanced contrastive views, effectively suppressing modality-specific noise during self-supervised learning. Second, an isomorphic graph integrating users and items is constructed to strengthen graph connectivity and mitigate the effects of data sparsity on graph representation learning. Finally, a joint optimization framework is designed to combine multitask learning with diffusion enhancement, enabling the collaborative optimization of recommendation objectives and enhancement signals while learning semantically consistent and generalizable multimodal representations. Experiments conducted on three real-world datasets, Baby, Sports, and Clothing, demonstrate that DEMGR consistently outperforms existing mainstream models in terms of Recall and normalized discounted cumulative gain(NDCG), achieving maximum improvements of 5.45% and 5.01%, respectively. These results verify the effectiveness of the proposed method for multimodal recommendation and provide valuable insights for the optimization of multimodal recommender systems.